Search

An unsupervised physics-informed neural operator-based wavefield back-propagation scheme is proposed, which enables the full wavefield information reconstruction along the link, nonlinearity compensation (0.46 dB Q-factor gain over 1StPS DBP), and fiber parameter identification.

Complete understanding of pedestrian behavior still poses challenges to current automated driving technologies. This study introduces a knowledge extraction framework to better understand the human perception of pedestrian crossing behavior from the ego-vehicle perspective. Using manually annotated explanatory texts, we train an information extraction algorithm to map the unstructured descriptive text into structured data. Evaluation of the algorithm confirms a good performance in most of the key feature categories. This model can identify and extract vital cues from pedestrian encounter descriptions, which can assist in analyzing human decision-making processes to estimate the pedestrians' situational crossing intent. The knowledge extracted from such structured text data can help to holistically study interconnection among the human-used factors which is crucial for modeling pedestrian behavior.

The electromigration (EM) of Cu/Sn-Ag1.8/Ni/Cu flip chip Cu pillar bumps structure under DC stress and bidirectional pulse current stress is studied. The temperature, thermal gradient and stress distribution of Cu pillar bumps under different current stress are obtained by ANSYS finite element simulation. Two groups of EM experiments under different current stress were carried out. At 125°C and ${3.5\times} {10^{4}\mathrm{A}/\text{cm}^{2}}$ , the resistance change rate of Cu pillar bumps under bidirectional pulse current stress was about 1/5 of the DC stress. After 220h of EM experiment, the bumps were completely alloyed into ${\text{Cu}_{6}\text{Sn}_{5}}$ and ${\text{Cu}_{3}\ \text{Sn}}$ under DC stress, and the polarity phenomenon appeared on the Ni layer and Cu pad at the cathode side. However, under the stress of bidirectional pulse current, the growth rate of intermetallic compound (IMC) is obviously slow, there is a small amount of solder remaining in the bump and the solder is biased to the side of the Ni layer, and the Ni layer does not show obvious polarity phenomenon. At the same time, combined with the simulation, it is found that the Cu trace may also become a weak link of the reliability of the Cu pillar bump.

A 32 Kb dual-port low-voltage SRAM in 28 nm FD-SOI, featuring foundry supplied high-density 6T bitcells, is presented. Dual-port configurability is realized by a unique dual-rail architecture, utilizing boost techniques that guarantee reliable operation in low-voltage. The area cost of the array is 62% lower, compared to widely used 8T two-port or dual-port SRAM arrays. The SRAM reliably operates in the low-voltage regime, and an access rate of 1MHz is measured at V MIN of 0.29 V. The highest energy efficiency of 1.35 fJ/bit-access is obtained at 80 MHz access rate, at a V DD of 0.54 V.

Topic modeling is a valuable method that enhances researchers’ ability to interpret information. Our research is the first to investigate neural topic modeling (NTM) methods for healthcare content analysis and topic discovery. We evaluated two NTM models to understand online information-seeking right after being diagnosed with breast cancer. Based on the achieved topic coherence and qualitative analysis on the topics discovered by each method, the NTM models work better than the traditional Latent Dirichlet Allocation (LDA) method when evaluated using the topic coherence score. The qualitative analysis shows that NTM models can find specific content-relevant topics, whereas LDA finds the general topics. Based on the results, we discover the main topics of online peerto-peer information seeking after the diagnosis of breast cancer, which included treatment and medication for different breast cancer types, symptom management after radiation treatments, insurance coverage of pathology, and family decision-making and support.

Hashing is widely-used in approximate nearest neighbor search for its computational efficiency. Most of the existing unsupervised hashing methods are based on local consistency that the Hamming distance between two images should be small if their features are similar. However, many false similar pairs may be included for the insufficient representation of features. Here we proposed deep unsupervised hashing by Global and Local Consistency (GLC). Specifically, GLC has two components named semantic information generating and semantic consistency learning, and each component is conducted from both global and local views. From local view, GLC introduces reliable graph and penalty graph to capture local signals with high confidence to preserve the semantic structure. From global view, GLC includes a distribution loss to capture the global consistency with cluster signals. Extensive experimental results on three widely-used benchmark datasets show that GLC performs better than existing state- of-the-art methods.

In recent years, the social web has been increasingly used for health information seeking, sharing, and subsequent health-related research. Women often use the Internet or social networking sites to seek information related to pregnancy in different stages. They may ask questions about birth control, trying to conceive, labor, or taking care of a newborn or baby. Classifying different types of questions about pregnancy information (e.g., before, during, and after pregnancy) can inform the design of social media and professional websites for pregnancy education and support. This research aims to investigate the attention mechanism built-in or added on top of the BERT model in classifying and annotating the pregnancy-related questions posted on a community Q&A site. We evaluated two BERT-based models and compared them against the traditional machine learning models for question classification. Most importantly, we investigated two attention mechanisms: the built-in self-attention mechanism of BERT and the additional attention layer on top of BERT for relevant term annotation. The classification performance showed that the BERT-based models worked better than the traditional models, and BERT with an additional attention layer can achieve higher overall precision than the basic BERT model. The results also showed that both attention mechanisms work differently on annotating relevant content, and they could serve as feature selection methods for text mining in general.

Interaction technology is one of the core technologies of augmented reality. With the maturity of augmented reality hardware technology and the expansion of application field, interaction technology has been paid more attention and development. This paper sorts out the classification methods of interaction technology, focuses on the research status of interaction technology based on computer vision, external devices and multi-modal, as well as the selection of 3D objects, introduces the latest typical application of interaction technology. Finally, the interaction technology is summarized and prospected.

This Research to Practice Full Paper presents a framework to evaluate and design flipped classroom activities for data science and management courses. Variants of flipped classrooms have been employed in STEM fields with great success in students’ learning outcomes. Research shows that flipped classrooms would improve students’ learning if it is implemented following rigorous procedures of an efficient instructional design. As a result, one of the critical focus of current flipped classroom research is what factors educators need to consider when designing a flipped learning environment. Currently, educators incorporate various factors such as "pre-recorded video lecture", "group activity" as a trial and error basis and adjust these factors based on their own experience and students’ feedback. On the other hand, the emergence of big data expects a new graduate to demonstrate mastery of concepts and skills for data acquisition, management, and analysis of inference from data when they enter the workforce. Currently, there is no systematic approach available to design a flipped classroom that is for the data science and management courses. In this research, we develop a framework first to investigate and evaluate the flipped classroom factors mentioned in the literature and identify a few that are most relevant to the two data management courses at our institute. Then, we classify each course topics into broader categories. So that the flipped classroom model can be developed for each category. For the flipped classroom for each category, we identify the pre-class and in-class activities to meet a certain learning objective for that topic category for each course. To evaluate the effectiveness of different factors as well as our flipped classroom models, students’ performance data, interviews, and surveys are conducted. This process is transformative and can be employed by other STEM disciplines to find the most influential factors to design effective flipped learning classrooms.

To make fully use of power transmission capacities of the tie-lines and reduce the total operating costs of the power grid, a marginal power generation cost based iterative optimization is proposed to optimize exchange power of the tie-lines. The proposed methodology iteratively optimizes exchange power of the tie-lines, starting from initial operation states of each regional power grid and initial exchange power of each tie-line without changing the results of unit commitment. The method can be easily applicated to engineering practice because it is closely related to the actual dispatching process of the power grid, and only requires information of the marginal unit operation in each regional power grid, protecting the data privacy of each regional power grid. Finally, a simulation is performed based on a case system of three regional grids and three tie-lines, and the simulation results verify the method.

In this research, we investigate and evaluate four unsupervised learning algorithms: K-Means and Self Organizing Maps (SOM), deep autoencoding Gaussian mixture model (DAGMM), and adversarially learned anomaly detection (ALAD) on two benchmark data sets towards network flow-based anomaly detection. We explore different parameters and neural network settings of the learning algorithms, respectively. The returned results show that DAGMM gains the lowest false-positive rate and a relatively high detection rate on one of the data set, whereas SOM achieves the best results on the other data set. By comparing the detection rates of all algorithms on attacks that have no instance in the training set, and we find that K-Means works the better than the other three. Whereas the comparison on the attacks that have few instances in the training set shows that the ALAD works much better than the others since it uses the adversarial sample generation mechanism. We conclude that to achieve the best performance, integration of the traditional and deep-learning algorithms for network flow-based intrusion detection is critical.

Research on social networks and understanding the interactions of the users can be modeled as a task of graph mining, such as predicting nodes and edges in networks. The challenges of building the graph representation include engineering features used by learning algorithms. Recent research in representation learning can automate the prediction by learning the features themselves. Many types of research have performed graph sampling using random walks or it’s derivatives. However, the random walk sometimes can not represent the features of the graph accurately enough. In this research, we propose BalNode2Vec – a new sampling algorithm for learning feature representations for nodes in networks by using balanced random walks. We define a notion of a nodes network neighborhood and design a balanced random walk procedure, which adapts to the graph topology. We show that through exploring the graph through a balanced random walk can generate richer representations. Efficacy of BalNode2vec over existing state-of-the-art techniques on link prediction is demonstrated by using several real-world networks from different domains.

Railway surface defect detection is an important technical measure to ensure the safe operation of the rail transit system. Due to the complex and diverse features of rail surface defects and the uneven curvature of the image caused by the track surface, it is difficult to obtain better detection results by traditional machine vision technology. The existing deep learning-based method only classifies the defect image and cannot locate the defect location. At the same time, the network structure of this method is complex, and the detection accuracy and speed cannot be well balanced. This paper proposes a new method for detecting railway surface defects based on end-to-end target detection and lightweight convolutional network structure. This method uses MobileNetV2 as the backbone network to extract image features, and then performs multi-scale defect detection to ensure real-time. Under the premise of detection, higher detection accuracy can be obtained.

A frequency modulation (FM) radio fuze electronic counter countermeasure (ECCM) scheme is proposed in this paper to encounter the DRFM (digital radio frequency memory) repeat jamming which has posed a serious threat to the FM fuze. This scheme consists of pulse diversity and antijamming signal processing techniques in fractional Fourier domain (FFD). The pulse diversity is achieved through slope varying chirp signals which has stationary bandwidth but varied pulse duration at each pulse repetition interval. The antijamming signal processing techniques in FFD are designed for two different situations. On one hand, a repeat jamming suppression method based on twice FRFTs and notch filter in FFD is presented, considering that the jamming signal and the target echo signal make up the fuze’s received signal together. On the other hand, if the jamming signal and the target echo signal exist in the fuze’s received signal at different times, a signal identification method is presented to distinguish the jamming signal and the target echo signal. The received signal’s FRFTs with different transformation angles have different amplitude and bandwidth performances, and these differences are utilized by this identification method. At last, formula derivations and simulation results validate this proposed ECCM scheme.

This research to practice full paper describes our cross-course knowledge integration approach, which uses a project-based learning environment. The structure and instructional pedagogies of some of our data management concentration courses in our department have not changed in a long time. Most of our data management courses cover theoretical knowledge without showing their practical applications. As a result, students don’t find those courses interesting enough. In addition, none of our data management courses require cross-course interactions, which is the current instructional pedagogical direction. With the goal of addressing these issues and improving student learning, faculty employed a cross-course knowledge integration model to teach the advanced database design course CIT 44400. We redesigned this advanced database course as part of our data management curriculum enhancement also to align with the trends of the IT industry and to enable students to correlate knowledge learned in various courses. In this redesign process, we integrated a project that aligns with industrial projects in this higher-level course curriculum, so that students can integrate and apply theoretical knowledge gained in lower-level courses through participating in a project-based learning approach, using current technology of the field. Evaluation results show crosscourse knowledge integration-based pedagogy gives students comprehensive knowledge that improves students’ performance and helps them link and apply knowledge learned in various courses.

The National Research Council’s report states that cross-sectional studies of multiple courses within a discipline, or all courses in a major, would enhance the understanding of how people learn the concepts, practices, and ways of thinking of science and engineering and the nature and development of expertise in a discipline. In science and engineering, everevolving technology and information make integrative abilities necessary and especially valuable. In this study, we investigated cross-curricular pedagogy, by engaging undergraduate students of two disciplines in collaboration on a common, contextconnected project, so that students are better prepared for solving interdisciplinary problems in career settings. We implemented cross-curricular pedagogy in a network security course and a big data analytics course. The era of big data enables datadriven malicious detection, and big data analytics techniques have been applied to analyzing network logs to reinforce information security and predict abnormal behaviors, so these domains overlap. We investigated two forms of cross-curricular activities: one was integrated instructional units, and the other was crosscurricular knowledge integration projects. The results show significant improvements in students confidence in solving crossdisciplinary problems and a much better understanding of data analytics and information security, as well as the connections between them. This project is the first to study the loose integration of two context-connected courses that are taught in parallel.

The transient behavior for four different types of TVS (non-snapback, snapback, spark gap, varistor) is modeled using the same modeling framework. By a 10 ns VF-TLP, the quasi-static I-V curve and the transient turn-on are captured and modeled in ADS. The models are applied in a SEED simulation to investigate the strengths and weaknesses of the modeling frame.

In the interest of improving the reproducibility of air discharge during IEC 61000-4-2 ESD testing, different methods are investigated that reduce the variation of the spark current during air discharge. As the variation is caused by the interplay of the statistical time lag and the approach speed, methods are investigated that provide initial charge carriers for the avalanche initiation at the beginning of sparking. These methods include ionizers, cold plasma, humidity, exposure of the electrodes to UV light, different electrode materials, and surface shapes. Most of these methods reduce the variations of the spark currents. However, only some can be implemented during testing. The strongest effect was achieved by using cold plasma and by one surface material found in a gasket.

With the rapid development of 3D printing technology and the continuous breakthrough of new material technologies, 3D printing has received more and more attention in the fields of industry, medical, sports, and education. In the cloud manufacturing environment, how to efficiently manage 3D printing models is a critical issue that needs to be solved urgently. The paper begins with a study of 3D printing model aggregation and retrieval mechanism based on this problem. Firstly, we devised a set of 3D printing model management framework with high scalability. Under this framework, a meta-model library and feature library are established to realize the aggregation of 3D printing models. Then, we developed a sketch-based 3D model retrieval method, which can help platform users to create and retrieve personalized 3D printing models easily. This study provides new ideas and a reference pattern for the model design and retrieval methods of existing commercial 3D printing platforms.

Natural language processing has been split into two sub-fields: semantic and syntactic analysis. Semantic processing deals with creating vector representations for each word. The goal is for the vectors to be dense and that relationships between words are made clear by their vectors. Word2Vec Skip-Gram model is one such semantic word embedding model that processes the words into vectors based on the co-occurrences of the words within a fixed context window. Syntactic processing, on the other hand, deals with gleaning textual information from the structure of the text. A popular method of syntactic processing is dependency parsing, where connections are explicitly drawn between different pairs of words in a phrase or clause, where one word modifies the other. Human beings make use of both semantic and syntactic information when processing text. As such, combining semantic and syntactic processing is important for medical text analysis. The dependencies between words define specific medical concepts, such as diseases and symptoms. The aim of this research paper is to investigate an altered word vector model—the SD Skip-Gram model which incorporates the dependencies between words. The SD Skip-Gram model is compared against the basic Skip-Gram model for analyzing common words and medical disease and symptom concepts. The returned results demonstrate that the SD Skip-Gram model can identify the relationships between the common words as well as the basic Skip-Gram model and perform better than the basic Skip-Gram model in identifying the related disease and symptom concepts.

In order to improve the aerospace network transmission performance considering the limitation of hardware resources and power consumption, a dynamic bandwidth scheduling method is proposed for burst traffic in FC-AE-1553 switching network. The feasibility of this scheduling mechanism is analyzed from the perspective of network throughput and burst traffic average delay based on simulation results.

In the process of cloud manufacturing simulation platform, the service agent drives the manufacturing service to form a self-organization network, in order to conduct transactions and cooperation spontaneously. Thus, the service network of service agent and the cooperation network of service agent are formed. In this paper, the models of two kinds of networks are created by theory of set pair analysis, and an evaluation method for two kinds of network models is proposed. Finally, the network evaluation method is validated by simulation data.

Considering a product service supply chain (PSSC) that consists of a manufacturer and an integrator of sales and after-sales service in which the demand of product is influenced by both the product quality and service level, we set up a decision optimization model and obtain the optimal equilibrium solutions under decentralized and centralized decision-making respectively. We further incorporate a two-way cost sharing and fixed transfer payment contract to coordinate the supply chain. The conclusions show that this coordination contract can coordinate the supply chain and achieve Pareto improvement when the contract parameters under certain conditions. The effect of wholesale price on the profit of both is the key to eliminate the double marginalization effect. We find that when the influence of service level on demand becomes greater, profits of both the manufacturer and the integrator become larger.

This paper explores an approach to identify the maximal cluster of hosts based on the proposed data fusion and clustering algorithms. The data fusion algorithm works with the cross-testing F1-measure matrices of three supervised machine learning algorithms to identify the similarity of the hosts with regards to intrusion detection. After identifying the hosts similarities, the clustering algorithm is developed to identify the maximal cluster. Through applying the experimental data set, we have identified a maximal cluster which consists of 5 hosts out of 16 hosts in a network. This maximal cluster identification based on the data fusion and machine learning algorithms can detect the similar anomaly behaviors generated by the same hacking mechanism to multiple machines in a network at the same time period. Furthermore, integrating the generated learning models from multiple machines or clusters could generate a robust detection model without time consuming training process based on all the network flows.