基于深度信念网络的列车故障音频诊断方法

铁路在交通运输行业有着举足轻重的地位，一旦列车发生故障将会导致严重的生命财产损失。由于列车发生故障的概率相对较低，因此难以捕获列车的故障样本。针对上述问题，提出了一种无监督学习的列车故障识别方法，通过检测列车音频信号来识别列车故障。该方法基于深度信念网络(DBN)，利用小波包分解提取检测信号的特征向量并将其作为DBN的输入，待网络充分训练后，由训练好的DBN识别当前列车的运行状况。现场监测实验结果表明，该方法能够在无监督的条件下有效识别列车故障，保障了列车的运行安全。     

基于深度学习的秀丽隐杆线虫显微图像分割方法

利用计算机实现自动、准确的秀丽隐杆线虫(C.elegans)的各项形态学参数分析,至关重要的是从显微图像上分割出线虫体态,但由于显微镜下的图像噪声较多,线虫边缘像素与周围环境相似,而且线虫的体态具有鞭毛和其他附着物需要分离,多方面因素导致设计一个鲁棒性的C.elegans分割算法仍然面临着挑战。针对这些问题,提出了一种基于深度学习的线虫分割方法,通过训练掩模区域卷积神经网络(Mask R-CNN)学习线虫形态特征实现自动分割。首先,通过改进多级特征池化将高级语义特征与低级边缘特征融合,结合大幅度软最大损失(LMSL)损失算法改进损失计算;然后,改进非极大值抑制;最后,引入全连接融合分支等方法对分割结果进行进一步优化。实验结果表明,相比原始的Mask R-CNN,该方法平均精确率(AP)提升了4.3个百分点,平均交并比(mIOU)提升了4个百分点。表明所提出的深度学习分割方法能够有效提高分割准确率,在显微图像中更加精确地分割出线虫体。     

Influence of pores arrangement on stability of photonic structures during sintering

Discrete Element Method (DEM) has been used for numerical investigation of sintering-induced structural deformations occurring in inverse opal photonic structures. The influence of the initial arrangement of template particles on the stability of highly porous inverse opal α-Al₂O₃ structures has been analyzed. The material transport, densification, as well as formation of defects and cracks have been compared for various case studies. Three different stages of defects formation have been distinguished starting with local defects ending with intrapore cracks. The results show that the packing of the template particles defined during the template self-assembly process play a crucial role in the later structural deformation upon thermal exposure. The simulation results are in very good agreement with experimental data obtained from SEM images and previous studies by ptychographic X-ray tomography.     

Damage identification in concrete using impact non-linear reverberation spectroscopy

High amplitude non-linear acoustic methods have shown potential for the identification of micro damage in brittle materials such as concrete. Commonly, these methods evaluate a non-linearity parameter from the relative change in frequency and attenuation with strain amplitude. Here, a novel attenuation model is introduced to describe the free reverberation from a standard impact resonance frequency test, together with an algorithm for estimating the unknown model coefficients. The non-linear variation can hereby by analyzed over a wider dynamic range as compared to conventional methods. The experimental measurement is simple and fully compatible with the standardized free-free linear impact frequency test.     

A new reverse reduce-error ensemble pruning algorithm

Although greedy algorithms possess high efficiency, they often receive suboptimal solutions of the ensemble pruning problem, since their exploration areas are limited in large extent. And another marked defect of almost all the currently existing ensemble pruning algorithms, including greedy ones, consists in: they simply abandon all of the classifiers which fail in the competition of ensemble selection, causing a considerable waste of useful resources and information. Inspired by these observations, an interesting greedy Reverse Reduce-Error (RRE) pruning algorithm incorporated with the operation of subtraction is proposed in this work. The RRE algorithm makes the best of the defeated candidate networks in a way that, the Worst Single Model (WSM) is chosen, and then, its votes are subtracted from the votes made by those selected components within the pruned ensemble. The reason is because, for most cases, the WSM might make mistakes in its estimation for the test samples. And, different from the classical RE, the near-optimal solution is produced based on the pruned error of all the available sequential subensembles. Besides, the backfitting step of RE algorithm is replaced with the selection step of a WSM in RRE. Moreover, the problem of ties might be solved more naturally with RRE. Finally, soft voting approach is employed in the testing to RRE algorithm. The performances of RE and RRE algorithms, and two baseline methods, i.e., the method which selects the Best Single Model (BSM) in the initial ensemble, and the method which retains all member networks of the initial ensemble (ALL), are evaluated on seven benchmark classification tasks under different initial ensemble setups. The results of the empirical investigation show the superiority of RRE over the other three ensemble pruning algorithms.     

THE effect of Srandom DOPANT fluctuation on threshold voltage and drain current variation in junctionless nanotransistors

We investigate the effect of dopant random fluctuation on threshold voltage and drain current variation in a two-gate nanoscale transistor. We used a quantum-corrected technology computer aided design simulation to run the simulation (10000 randomizations). With this simulation, we could study the effects of varying the dimensions (length and width), and thicknesses of oxide and dopant factors of a transistor on the threshold voltage and drain current in subthreshold region (off) and overthreshold (on). It was found that in the subthreshold region the variability of the drain current and threshold voltage is relatively fixed while in the overthreshold region the variability of the threshold voltage and drain current decreases remarkably, despite the slight reduction of gate voltage diffusion (compared with that of the subthreshold). These results have been interpreted by using previously reported models for threshold current variability, load displacement, and simple analytical calculations. Scaling analysis shows that the variability of the characteristics of this semiconductor increases as the effects of the short channel increases. Therefore, with a slight increase of length and a reduction of width, oxide thickness, and dopant factor, we could correct the effect of the short channel.     

Evolutionary collective behavior decomposition model for time series data mining

In this research, we propose a novel framework referred to as collective game behavior decomposition where complex collective behavior is assumed to be generated by aggregation of several groups of agents following different strategies and complexity emerges from collaboration and competition of individuals. The strategy of an agent is modeled by certain simple game theory models with limited information. Genetic algorithms are used to obtain the optimal collective behavior decomposition based on history data. The trained model can be used for collective behavior prediction. For modeling individual behavior, two simple games, the minority game and mixed game are investigated in experiments on the real-world stock prices and foreign-exchange rate. Experimental results are presented to show the effectiveness of the new proposed model.     

Fully coupled LES-DEM of particle interaction and agglomeration in a turbulent channel flow

Coupled large eddy simulation and the discrete element method are applied to study turbulent particle–laden flows, including particle dispersion and agglomeration, in a channel. The particle–particle interaction model is based on the Hertz–Mindlin approach with Johnson–Kendall–Roberts cohesion to allow the simulation of van der Waals forces in a dry air flow. The influence of different particle surface energies, and the impact of fluid turbulence, on agglomeration behaviour are investigated. The agglomeration rate is found to be strongly influenced by the particle surface energy, with a positive relationship observed between the two. Particle agglomeration is found to be enhanced in two separate regions within the channel. First, in the near-wall region due to the high particle concentration there driven by turbophoresis, and secondly in the buffer region where the high turbulence intensity enhances particle–particle interactions.     

Fault handling in large water networks with online dictionary learning

Fault detection and isolation in water distribution networks is an active topic due to the nonlinearities of flow propagation and recent increases in data availability due to sensor deployment. Here, we propose an efficient two-step data driven alternative: first, we perform sensor placement taking the network topology into account; second, we use incoming sensor data to build a network model through online dictionary learning. Online learning is fast and allows tackling large networks as it processes small batches of signals at a time. This brings the benefit of continuous integration of new data into the existing network model, either in the beginning for training or in production when new data samples are gathered. The proposed algorithms show good performance in our simulations on both small and large-scale networks.