56016 research outputs found
Sort by
Effect of thermo-mechanical activation of waste concrete powder (WCP) on the characteristics of cement mixtures
No full textThe objective of present study is to utilize and study the reactivity and hydration behaviour of thermo-mechanically activated waste concrete powder (WCP). For activation, the WCP was burned for 2 h at 650 ??C, followed by mechanical grinding through ball milling. The reactivity of the activated WCP was improved significantly, as corroborated through the Chapelle test, hydration analysis and compressive strength analysis. Different blends were prepared with 10 to 40 % OPC replacement. Activated WCP was found to be effective in enhancing the reactivity of cement mixtures for all replacement proportions and developed comparable compressive strength in comparison to OPC. At 56 days, the application of activated WCP resulted in the developed of 75.5 % better compressive strength than the application of as received WCP. Mineralogical analysis exhibited the development of more secondary C-S-H gel with the formation of calcite at the later age. Using lifecycle analysis software, the replacement of 40 wt% of OPC with activated WCP was calculated with 37.6 %, 39.8 % and 19.7 % less CO2 emissions, less mineral resources consumption and less energy consumption, respectively
Effects of unlimited angular motion cue and cue discrepancy on simulator sickness
No full textSimulator sickness is a crucial concern undermining several benefits of simulator training, such as a realistic environment, low costs, and safe practice of emergencies. This study investigated the effects of unbounded angular motions and visual-vestibular cue discrepancies on simulator sickness for flight simulator training. Human subject experiments with 36 participants demonstrated that simulator sickness, measured by question-naires and physiological signals, was significantly decreased by offering both motion and visual cues rather than visual signals alone (p < 0.05). Specifically, nausea (without motion = 54.59, with motion = 31.27; p = 0.036) and disorientation scores (without motion = 81.20, with motion = 44.08; p = 0.028) significantly decreased when both motion and visual signals were present. Furthermore, the experimental results showed a significant correlation between simulator sickness and visual-vestibular cue mismatches, particularly for the angular ve-locity along the z-axis (r = 0.110, p = 0.04). The pitch angle discrepancy (r = 0.156, p = 0.004) between the visual and motion cues was significantly correlated with the sickness severity, unlike the roll angle disparity (r =-0.009, p = 0.871). The results from this study can be explored for flight training operations using motion simulators to minimize or eliminate simulator sickness
Nanoscale Etching of La0.7Sr0.3MnO3 without Etch Lag using Chlorine Based Inductively Coupled Plasma
No full textLa0.7Sr0.3MnO3 (LSMO) has been considered as a promising material for future electronic and spintronic device application due to its unique properties such as pure spin polarization, colossal magnetoresistance, and high temperature coefficient of resistance (TCR). To apply this promising material for practical application, large epitaxial LSMO layers should be etched into micro- and nano-scale device structures. However, a comprehensive study on the etch of LSMO has not been demonstrated yet. Herein, the etch rates of LSMO are studied using inductively coupled plasma reactive ion etching (ICP-RIE) method, while controlling critical etching parameters such as ICP source power, radio frequency (rf) chuck power, etching gas ratio, and chamber pressure. We found that the etching process can be applied to nanoscale structures (down to 100 nm) without etch lag effect, exhibiting smaller etch depth in smaller features. This study will provide a good reference for the etching and the engineering of LSMO toward future electronic and spintronic devices such as highly sensitive bolometers and low-power memory devices
A Comprehensive Study of Deep Learning for Real-World Multivariate Time Series Classification
No full textDepartment of Industrial Engineeringope
Bytee : The Interactive System Based on Visual Imagery to Encourage Personal Yoga Practices
No full textDepartment of DesignInterest in in-home workouts and yoga has been steadily increasing, and various interactive services have emerged to provide new experiences for novice yoga trainees to do yoga alone at home. However, it is difficult for trainees themselves to do yoga constantly, and it is also insufficient to induce motivation for exercise. To address the difficulty of doing yoga consistently, I have developed a system called Bytee, which uses visual imagery and pose estimation based on AI to encourage yoga practice for novice trainees. The system consists of three features, based on visual imagery and an AI-based framework. The three-week experiment was performed by 12 participants using the Bytee system and another 11 participants using video streaming. The change in motivation was analyzed through three survey sessions using a customized questionnaire based on the ARCS motivational model clustered into four categories: attention (A), relevance (R), confidence (C), and satisfaction (S). The detailed features that influenced motivation were identified through semi-structured interviews. This study derived three insights through quantitative and qualitative analysis: 1) challenges and motivational factors for workouts from the novice yoga trainees' perspectives when practicing alone at home2) the system design consisting of three visual features with different design properties and an AI-based frameworkand 3) findings on how this system impacts novice trainees to encourage yoga based on the ARCS motivational model. I discuss the key implications of a system, composed of visual imagery and an AI-based framework to encourage yoga practice.clos
????????? ???????????? ????????? 300 GHz ??????????????? ????????? ??????
No full textDepartment of Physicsclos
Direct observation of the hydrogen permeation in the hydrogen storage alloys
No full textDepartment of Materials Science and Engineeringclos
Enabling Deep Neural Network Inferences on Resource-constraint Devices
Get PDFDepartment of Computer Science and EngineeringWhile deep neural networks (DNN) are widely used on various devices, including resource-constraint devices such as IoT, AR/VR, and mobile devices, running DNN from resource-constrained devices remains challenging. There exist three approaches for DNN inferences on resource-constraint devices: 1) lightweight DNN for on-device computing, 2) offloading DNN inferences to a cloud server, and 3) split computing to utilize computation and network resources efficiently.
Designing a lightweight DNN without compromising the accuracy of DNN is challenging due to a trade-off between latency and accuracy, that more computation is required to achieve higher accuracy. One solution to overcome this challenge is pre-processing to extract and transfer helpful information to achieve high accuracy of DNN. We design the pre-processing, which consists of three processes. The first process of pre-processing is finding out the best input source. The second process is the input-processing which extracts and contains important information for DNN inferences among the whole information gained from the input source. The last process is choosing or designing a suitable lightweight DNN for processed input. As an instance of how to apply the pre-processing, in Sec 2, we present a new transportation mode recognition system for smartphones called DeepVehicleSense, which aims at achieving three performance objectives: high accuracy, low latency, and low power consumption at once by exploiting sound characteristics captured from the built-in microphone while being on candidate transportations. To achieve high accuracy and low latency, DeepVehicleSense makes use of non-linear filters that can best extract the transportation sound samples. For the recognition of five different transportation modes, we design a deep learning-based sound classifier using a novel deep neural network architecture with multiple branches. Our staged inference technique can significantly reduce runtime and energy consumption while maintaining high accuracy for the majority of samples.
Offloading DNN inferences to a server is a solution for DNN inferences on resource-constraint devices, but there is one concern about latency caused by data transmission. To reduce transmission latency, recent studies have tried to make this offloading process more efficient by compressing data to be offloaded. However, conventional compression techniques are designed for human beings, so they compress data to be possible to restore data, which looks like the original from the perspective of human eyes. As a result, the compressed data through the compression technique contains redundancy beyond the necessary information for DNN inference.
In other words, the most fundamental question on extracting and offloading the minimal amount of necessary information that does not degrade the inference accuracy has remained unanswered. To answer the question, in Sec 3, we call such an ideal offloading semantic offloading and propose N-epitomizer, a new offloading framework that enables semantic offloading, thus achieving more reliable and timely inferences in highly-fluctuated or even low-bandwidth wireless networks. To realize N-epitomizer, we design an autoencoder-based scalable encoder trained to extract the most informative data and scale its output size to meet the latency and accuracy requirements of inferences over a network.
Even though our proposed lightweight DNN and offloading framework with the essential information extractor achieve low latency while preserving DNN performance, they alone cannot realize latency-guaranteed DNN inferences. To realize latency-guaranteed DNN inferences, the computational complexity of the lightweight DNN and the compression performance of the encoder for offloading should be adaptively selected according to current computation resources and network conditions by utilizing the DNN's trade-off between computational complexity and DNN performance and the encoder's trade-off between compression performance and DNN performance. To this end, we propose a new framework for latency-guaranteed DNN inferences called LG-DI, which predicts DNN performance degradation given a latency budget in advance and utilizes the better method between the lightweight DNN and offloading with compression. As a result, our proposed framework for DNN inferences can guarantee latency regardless of changes in computation and network resources while maintaining DNN performance as much as possible.ope
Enhanced Four-Port Dual-Active-Bridge Converter Employing Power Decoupling Capability for DC Microgrid system
No full textDepartment of Electrical Engineeringclos
Fast Monocular Depth Estimation with Neural Network Compression for Collision Avoidance
No full textDepartment of Mechanical EngineeringThis paper introduces monocular depth estimation leveraging neural network compression algorithms which make a deep neural network smaller and faster for embedded systems and a collision avoidance algorithm with deep reinforcement learning. In this study, we employ an auto-encoder architecture for depth estimation, with ResNet-50 as the encoder and multiple convolutional layers as the decoder. As the network is too heavy to operate on embedded systems in real-time, knowledge distillation and neural network quantization and tensor decomposition, especially tucker-2 decomposition, are employed to reduce the number of parameters and latency. We employ knowledge distillation to initialize the weights of our depth estimator to improve the accuracy of depth estimation. We use tucker-2 decomposition to reduce the number of parameters and variational bayesian matrix factorization to estimate the ranks for the tucker decomposition, yielding multiple tensors that approximate the original weight. As a result of the decomposition, the accuracy of the network may degenerate, so we fine-tune the neural network to recover the accuracy. We also apply neural network quantization which converts computational structure from floating point representation to fixed point representation to accelerate the neural network on embedded systems. As a result of the neural network compression, an embedded computer board can execute the deep neural network faster with less memory footprint while maintaining reasonable accuracy. We integrate the network into a collision avoidance algorithm based on deep reinforcement learning whose states are depth information from our depth estimator. We demonstrate that our depth estimator can operate at 17 frames per second on NVIDIA Jetson TX2 CPU and that the agent can avoid obstacles at a high speed.ope