自然场景车标数据集的构建及其应用

车标作为车辆身份的关键特征之一,在车辆的监控与辨识中发挥着重要作用。由于自然场景复杂多变,对其中的车标进行准确识别仍具有很大的挑战性。目前公开数据库很少且存在诸多局限,导致研究缺乏可信度和实用性。本文建立了一个面向自然场景的全新数据集,包含多种采集环境下的10 324幅、67类车辆图像。基于此数据集开展应用研究,提出一个目标检测与深度学习相结合的车标识别方法,包括车标区域定位和车标种类预测两大步骤。实验表明,该方法对复杂背景有较强的适应性,在涉及30种车标的分类任务中达到89.0%的总体识别率。     

Predicting intensive care unit bed occupancy for integrated operating room scheduling via neural networks

In a master surgery scheduling (MSS) problem, a hospital's operating room (OR) capacity is assigned to different medical specialties. This task is critical since the risk of assigning too much or too little OR time to a specialty is associated with overtime or deficit hours of the staff, deferral or delay of surgeries, and unsatisfied—or even endangered—patients. Most MSS approaches in the literature focus only on the OR while neglecting the impact on downstream units or reflect a simplified version of the real‐world situation. We present the first prediction model for the integrated OR scheduling problem based on machine learning. Our three‐step approach focuses on the intensive care unit (ICU) and reflects elective and urgent patients, inpatients and outpatients, and all possible paths through the hospital. We provide an empirical evaluation of our method with surgery data for Universitätsklinikum Augsburg, a German tertiary care hospital with 1700 beds. We show that our model outperforms a state‐of‐the‐art model by 43% in number of predicted beds. Our model can be used as supporting tool for hospital managers or incorporated in an optimization model. Eventually, we provide guidance to support hospital managers in scheduling surgeries more efficiently.     

Maneuvering target tracking of UAV based on MN-DDPG and transfer learning

Tracking maneuvering target in real time autonomously and accurately in an uncertain environment is one of the challenging missions for unmanned aerial vehicles(UAVs).In this paper,aiming to address the control problem of maneuvering target tracking and obstacle avoidance,an online path planning approach for UAV is developed based on deep reinforcement learning.Through end-to-end learning powered by neural networks,the proposed approach can achieve the perception of the environment and continuous motion output control.This proposed approach includes:(1)A deep deterministic policy gradient(DDPG)-based control framework to provide learning and autonomous decision-making capa-bility for UAVs;(2)An improved method named MN-DDPG for introducing a type of mixed noises to assist UAV with exploring stochastic strategies for online optimal planning;and(3)An algorithm of task-decomposition and pre-training for efficient transfer learning to improve the generalization capability of UAV's control model built based on MN-DDPG.The experimental simulation results have verified that the proposed approach can achieve good self-adaptive adjustment of UAV's flight attitude in the tasks of maneuvering target tracking with a significant improvement in generalization capability and training efficiency of UAV tracking controller in uncertain environments.     

Scheduling maintenance and semiresumable jobs on a single machine

The majority of scheduling literature assumes that the machines are available at all times. In this paper, we study single machine scheduling problems where the machine maintenance must be performed within certain intervals and hence the machine is not available during the maintenance periods. We also assume that if a job is not processed to completion before the machine is stopped for maintenance, an additional setup is necessary when the processing is resumed. Our purpose is to schedule the maintenance and jobs to minimize some performance measures. The objective functions that we consider are minimizing the total weighted job completion times and minimizing the maximum lateness. In both cases, maintenance must be performed within a fixed period T, and the time for the maintenance is a decision variable. In this paper, we study two scenarios concerning the planning horizon. First, we show that, when the planning horizon is long in relation to T, the problem with either objective function is NP-complete, and we present pseudopolynomial time dynamic programming algorithms for both objective functions. In the second scenario, the planning horizon is short in relation to T. However, part of the period T may have elapsed before we schedule any jobs in this planning horizon, and the remaining time before the maintenance is shorter than the current planning horizon. Hence we must schedule one maintenance in this planning horizon. We show that the problem of minimizing the total weighted completion times in this scenario is NP-complete, while the shortest processing time (SPT) rule and the earliest due date (EDD) rule are optimal for the total completion time problem and the maximum lateness problem respectively. © 1999 John Wiley & Sons, Inc. Naval Research Logistics 46: 845–863, 1999     

模型不可知的联合相互学习

主流的联邦学习(federated learning, FL)方法需要梯度的交互和数据同分布的理想假定,这就带来了额外的通信开销、隐私泄露和数据低效性的问题。因此,提出了一种新的FL框架,称为模型不可知的联合相互学习 (model agnostic federated mutual learning, MAFML)。MAFML仅利用少量低维的信息(例如,图像分类任务中神经网络输出的软标签)共享实现跨机构间的“互学互教”,且MAFML不需要共享一个全局模型,机构用户可以自定制私有模型。同时,MAFML使用简洁的梯度冲突避免方法使每个参与者在不降低自身域数据性能的前提下,能够很好地泛化到其他域的数据。在多个跨域数据集上的实验表明,MAFML可以为面临“竞争与合作”困境的联盟企业提供一种有前景的解决方法。     

Allocation of quality improvement targets based on investments in learning

Purchased materials often account for more than 50% of a manufacturer's product nonconformance cost. A common strategy for reducing such costs is to allocate periodic quality improvement targets to suppliers of such materials. Improvement target allocations are often accomplished via ad hoc methods such as prescribing a fixed, across‐the‐board percentage improvement for all suppliers, which, however, may not be the most effective or efficient approach for allocating improvement targets. We propose a formal modeling and optimization approach for assessing quality improvement targets for suppliers, based on process variance reduction. In our models, a manufacturer has multiple product performance measures that are linear functions of a common set of design variables (factors), each of which is an output from an independent supplier's process. We assume that a manufacturer's quality improvement is a result of reductions in supplier process variances, obtained through learning and experience, which require appropriate investments by both the manufacturer and suppliers. Three learning investment (cost) models for achieving a given learning rate are used to determine the allocations that minimize expected costs for both the supplier and manufacturer and to assess the sensitivity of investment in learning on the allocation of quality improvement targets. Solutions for determining optimal learning rates, and concomitant quality improvement targets are derived for each learning investment function. We also account for the risk that a supplier may not achieve a targeted learning rate for quality improvements. An extensive computational study is conducted to investigate the differences between optimal variance allocations and a fixed percentage allocation. These differences are examined with respect to (i) variance improvement targets and (ii) total expected cost. For certain types of learning investment models, the results suggest that orders of magnitude differences in variance allocations and expected total costs occur between optimal allocations and those arrived at via the commonly used rule of fixed percentage allocations. However, for learning investments characterized by a quadratic function, there is surprisingly close agreement with an “across‐the‐board” allocation of 20% quality improvement targets. © John Wiley & Sons, Inc. Naval Research Logistics 48: 684–709, 2001     

局部放电在线监测电流传感器的等效电路及其参数计算

局部放电在线监测是保证电力设备正常工作的重要手段。从变压器理论出发 ,提出了局放脉冲电流传感器等效电路及其参数计算方法 ,所提出的方法已成功地应用于实际局放在线监测传感器的设计中     

一种基于局部不变特征的图像特定场景检测方法

图像场景的自动检测,对于图像的标注以及语义检索具有非常重要的作用。本文研究根据实际应用的需要,围绕会晤、集会、海滩等八类特定场景图像的检测问题展开。首先对图像进行局部关键点的检测以及SIFT特征描述子的计算,从而提取图像的局部特征,在此基础上基于支撑向量机构建多分类器,进行特征训练,最终获得较为准确的检测结果。实验重点针对分类器核函数的确定以及特征选取策略等问题展开,实验结果表明,采用径向基核函数构建多分类器以及特征点按尺度大小排序取前n位的选取策略可以获得较为准确和鲁棒的特定场景检测结果。本方法在保证满足一定程度场景检测准确率的前提下,具有简单快速的特点,能够满足实际应用的需要。     

基于VMD和SVM的舰船辐射噪声特征提取及分类识别

针对复杂海洋背景下舰船声频辐射噪声特征提取困难的问题,提出一种基于变分模态分解、中心频率、复杂度特征和支持向量机的舰船辐射噪声特征提取及分类识别方法。对四类舰船辐射噪声信号使用变分模态方法分解,得到一定数量的固有模态函数。通过比较提取能量最大的固有模态函数中心频率和排列熵作为特征参数,并利用支持向量机方法对四类舰船信号样本进行分类识别。实验结果表明,该方法可以实现对舰船辐射噪声的特征提取,与已有方法对比,该方法具有较高的识别率。     

基于深度学习的MPSK信号调制识别

为了有效实现信号调制方式的智能识别,提出基于深度学习的多进制相移键控(Multiple Phase Shift Keying, MPSK)信号调制识别方法。分析接收MPSK信号的循环谱,并通过提取MPSK信号循环谱的等高图获得二维特征信息,利用深度学习中的卷积神经网络对二维特征进行训练,使用测试样本对所设计的调制识别方法的有效性进行验证。仿真结果表明,所提方法具有良好的识别性能。