|
主题:Frontier Machine Intelligence Technology and Applications
时间:2022年9月28日13:00-16:30
#腾讯会议:982-400-847
主持人: 罗辛 研究员 中国科学院重庆绿色智能技术研究院
时间 | 报告人 | 报告题目 |
13:00-13:30 | 李响 副教授 西安交通大学 | 基于深度学习的机械装备剩余寿命智能预测方法 |
13:30-14:15 | 王锐 副研究员 国防科技大学 | 组合优化:进化计算与深度强化学习 |
14:15-15:00 | 叶茂娇 教授 南京理工大学 | Distributed Robust Nash Equilibrium Seeking for Games with Partial Decision Information |
15:00-15:45 | 王建坤 助理教授 南方科技大学 | Learning-based Sampling Strategies for Robot Path Planning |
15:45-16:30 | 陈正华 研究员 A*STAR, Singapore | Position Encoding Based Convolutional Neural Networks for Machine Remaining Useful Life Prediction |
报告人:李响 副教授,西安交通大学
时间:13:00-13:30
题目:基于深度学习的机械装备剩余寿命智能预测方法
摘要: 近年来,基于深度学习的机械装备剩余寿命预测方法得到了显著的发展。本报告首先介绍大数据驱动的机械装备寿命预测整体流程,然后主要介绍课题组在此方向的最新进展,包括传感器故障下的智能预测、跨领域迁移智能预测等前沿问题,以及循环一致性理论、正则化表征学习、深度领域对抗学习等最新方法,最后介绍当前存在的挑战与未来发展的方向。
报告人:王锐 副研究员,国防科技大学
时间: 13:30-14:15
题目:组合优化:进化计算与深度强化学习
摘要:组合优化问题广泛存在于国防、交通、工业、生活等各个领域,几十年来,经典运筹优化、启发式及群体智能(进化计算)方法是解决组合优化问题的主要手段,但随着实际应用中问题规模的不断扩大,求解实时性的要求越来越高,经典算法面临着很大的计算压力,很难实现组合优化问题的快速(在线)求解。近年来随着深度学习技术的迅猛发展,深度强化学习在围棋、机器人等领域的瞩目成果显示了其强大的学习能力与序贯决策能力,为组合优化问题的求解提供了一种新思路。本报告将介绍近些年利用深度强化学习方法解决组合优化问题的相关理论方法与应用研究,并探讨未来该方向亟待解决的若干问题。
报告人:叶茂娇 教授,南京理工大学
时间:14:15-15:00
题目:Distributed Robust Nash Equilibrium Seeking for Games with Partial Decision Information
摘要: Game theory has been widely adopted to deal with interactive decision-making problems in various distributed systems such as smart grids, wireless communication networks, traffic systems, etc. As a result, distributed Nash equilibrium seeking strategies have recently attracted increasing attention due to their remarkable advantages in relaxing the requirement of a central node for information broadcasting or full observation over players’ actions. This talk aims to provide a glimpse of recent advances on distributed Nash equilibrium seeking in games with partial decision information. In particular, some robust strategies will be discussed in details. Finally, some promising directions are suggested for future research.
报告人:王建坤 助理教授,南方科技大学
时间:15:00-15:45
题目:Learning-based Sampling Strategies for Robot Path Planning
摘要:Sampling-based path planning algorithms have become very popular due to their capability of efficiently searching the state space. However, the uniform sampling strategy cannot help the path planner quickly find an initial solution or converge to the optimal solution. Learning-based strategies can help achieve more efficient search by learning the construction information of the current environment. Fed by a number of successful path planning cases, the designed neural network can learn how to generate promising region in given environments. Then, promising region is used to guide the path planner to heuristically search the regions where a feasible solution probably exists, resulting in a nonuniform sampling strategy. The experiment results demonstrate that the path planning performance improves significantly. Meanwhile, we will investigate different learning-based sampling strategies and compare their performance on path planning problems.
报告人:陈正华 研究员,Institute for Infocomm Research, A*STAR, Singapore
时间:15:45-16:30
题目:Position Encoding Based Convolutional Neural Networks for Machine Remaining Useful Life Prediction
摘要:Accurate remaining useful life (RUL) prediction is vital for smart manufacturing, which helps to prevent machine failures and ensure system reliability. Recently, many deep learning based methods have been proposed to predict RUL. Among these methods, recurrent neural network (RNN) based approaches show a strong capability of capturing sequential information. This allows RNN based methods to perform better than convolutional neural network (CNN) based approaches on the RUL prediction task. We question this common paradigm and argue that existing CNN based approaches in the literature are not well-designed according to the classic principles of CNN, which hinders their performances. Additionally, the capacity of capturing sequential information is highly affected by the receptive field of CNN, which has been neglected by existing CNN based methods. To solve these problems, we present a series of new CNNs, which show competitive results to RNN based methods. To further improve the performance via well capturing the temporal information, a position encoding scheme is developed to enhance the sequential information encoded by a CNN. Hence, our proposed position encoding based CNN called PE-Net can be further improved and even performs better than RNN based methods. Extensive experiments are conducted on the C-MAPSS dataset, where our PE-Net shows state-of-the-art performance.
Archiver|手机版|科学网 ( 京ICP备07017567号-12 )
GMT+8, 2024-12-23 09:30
Powered by ScienceNet.cn
Copyright © 2007- 中国科学报社