关键词： Laminar flame speed   Machine learning   Reduced mechanism  

摘要： The focus of the present work is to investigate a new methodology for the rapid generation of laminar flame speed lookup tables, to be used replacing correlation laws in internal combustion engine simulations. Current production engines run mostly under the thickened wrinkled flame combustion regime, which allows the application of a flamelet modelling approach, which requires the a-priori evaluation of the laminar flame speed and thickness. The use of correlation laws, derived from experimental data, has the advantage to be extremely fast to compute, but displays a lack of precision in conditions far from the experimental reference, as it usually happens for ICE applications. On the other hand, the detailed chemical simulation of a freely propagating adiabatic flame, performed for a sufficiently refined grid of reference points, to be interpolated during runtime, might require hundreds of hours of computing. The use of a reduced chemical mechanism can potentially cut by orders of magnitude the required time, but on the other hand it will decrease accuracy. In the present work, the potential of integrating machine learning algorithms and neural networks in the workflow with different approaches was valuated, leveraging the potential of new and optimized software libraries, to reduce simulation times while maintaining a high level of accuracy, with respect to the results obtained with the complete scheme. (C) 2020 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

关键词： automatic license plate recognition   intelligent vehicle access   histogram of oriented gradients   artificial neural networks  

摘要： Automatic vehicle license plate recognition is an essential part of intelligent vehicle access control and monitoring systems. With the increasing number of vehicles, it is important that an effective real-time system for automated license plate recognition is developed. Computer vision techniques are typically used for this task. However, it remains a challenging problem, as both high accuracy and low processing time are required in such a system. Here, we propose a method for license plate recognition that seeks to find a balance between these two requirements. The proposed method consists of two stages: detection and recognition. In the detection stage, the image is processed so that a region of interest is identified. In the recognition stage, features are extracted from the region of interest using the histogram of oriented gradients method. These features are then used to train an artificial neural network to identify characters in the license plate. Experimental results show that the proposed method achieves a high level of accuracy as well as low processing time when compared to existing methods, indicating that it is suitable for real-time applications.

关键词： Human activity recognition   Image processing   Speeded up robust features   Bag of visual words   Machine learning   k-Nearest neighbors   Decision tree   Support vector machine   Naive Bayes   Hyperparameter optimization  

摘要： Human activity recognition (HAR) has quite a wide range of applications. Due to its widespread use, new studies have been developed to improve the HAR performance. In this study, HAR is carried out using the commonly preferred KTH and Weizmann dataset, as well as a dataset which we created. Speeded up robust features (SURF) are used to extract features from these datasets. These features are reinforced with bag of visual words (BoVW). Different from the studies in the literature that use similar methods, SURF descriptors are extracted from binary images as well as grayscale images. Moreover, four different machine learning (ML) methods such as k-nearest neighbors, decision tree, support vector machine and naive Bayes are used for classification of BoVW features. Hyperparameter optimization is used to set the hyperparameters of these ML methods. As a result, ML methods are compared with each other through a comparison with the activity recognition performances of binary and grayscale image features. The results show that if the contrast of the environment decreases when a human enters the frame, the SURF of the binary image are more effective than the SURF of the gray image for HAR.

关键词： UAV   visual RGB   real-time   YOLOv3   color detection   object detection   machine learning system  

摘要： What makes unmanned aerial vehicles (UAVs) intelligent is their capability of sensing and understanding new unknown environments. Some studies utilize computer vision algorithms like Visual Simultaneous Localization and Mapping (VSLAM) and Visual Odometry (VO) to sense the environment for pose estimation, obstacles avoidance and visual servoing. However, understanding the new environment (i.e., make the UAV recognize generic objects) is still an essential scientific problem that lacks a solution. Therefore, this paper takes a step to understand the items in an unknown environment. The aim of this research is to enable the UAV with basic understanding capability for a high-level UAV flock application in the future. Specially, firstly, the proposed understanding method combines machine learning and traditional algorithm to understand the unknown environment through RGB images;secondly, the You Only Look Once (YOLO) object detection system is integrated (based on TensorFlow) in a smartphone to perceive the position and category of 80 classes of objects in the images;thirdly, the method makes the UAV more intelligent and liberates the operator from labor;fourthly, detection accuracy and latency in working condition are quantitatively evaluated, and properties of generality (can be used in various platforms), transportability (easily deployed from one platform to another) and scalability (easily updated and maintained) for UAV flocks are qualitatively discussed. The experiments suggest that the method has enough accuracy to recognize various objects with high computational speed, and excellent properties of generality, transportability and scalability.

关键词： chemical logistic warehouses   hazard identification   hazard index   machine learning clustering  

摘要： With the rapid development of chemical process plants worldwide, the safe storage of hazardous chemicals continues to be an important topic. Chemical warehouse incidents related to fire and explosion are reported constantly. These incidents not only cause direct economic loss but also harm the local environment. Therefore, an accurate hazard identification method for a logistic warehouse is valuable since it can help people recognize the potential hazard in the chemical storage area. In this study, a storage hazard factor (SHF) was developed to evaluate and rank the inherent hazards of chemicals stored in logistic warehouses. In the calculation of SHF, the National Fire Protection Association (NFPA) hazard diamond was used to quantify the instability hazard;the U.S. Department of Energy's Protective Action Criteria (PAC) values were utilized to modify chemical health hazard, and two machine learning based methods were used for the classification of flammability. The SHF and safety penalty factors (e.g., quantity, population density) were then utilized to create a hazard index of the facilities. This index can be used not only for developing a proper emergency response plan but also for effectively communicating risk with the residents who live near the facility. In addition, the index can be used by governments to allocate the resources better for first responders to make fire protection strategies, and by stakeholders to improve the risk management.

关键词： Machine learning   Mass flux   Species segregation   Voidage   Cluster   Circulating fluidized bed  

摘要： Machine learning methods were applied to circulating fluidized bed (CFB) riser data. The goals were to (i) provide insights on various fluidization phenomena through determining the relative dominance of the process variables, and (ii) develop a model to provide predictive capability in the absence of firstprinciples understanding that remains elusive. The Random Forest results indicate radial position had the most dominant influence on local mass flux and species segregation, overall mass flux was the most dominant for local particle concentration, while no variable was particularly dominant or negligible for the local clustering characteristics. Furthermore, the Neural Network can be trained to provide good predictive capability, without any mechanistic understanding needed, if a sufficiently large dataset is used for training and if the input variables fully account for all the effects at play. This study underscores the value of machine learning methods in fluidization to advance understanding and provide adequate predictions. (C) 2020 Elsevier Ltd. All rights reserved.

关键词： Large-scale optimization   non-convex   machine learning   trust-region methods   quasi-Newton methods   limited-memory symmetric rank-one update  

摘要： Machine learning (ML) problems are often posed as highly nonlinear and nonconvex unconstrained optimization problems. Methods for solving ML problems based on stochastic gradient descent are easily scaled for very large problems but may involve fine-tuning many hyper-parameters. Quasi-Newton approaches based on the limited-memory Broyden-Fletcher-Goldfarb-Shanno (BFGS) update typically do not require manually tuning hyper-parameters but suffer from approximating a potentially indefinite Hessian with a positive-definite matrix. Hessian-free methods leverage the ability to perform Hessian-vector multiplication without needing the entire Hessian matrix, but each iteration's complexity is significantly greater than quasi-Newton methods. In this paper we propose an alternative approach for solving ML problems based on a quasi-Newton trust-region framework for solving large-scale optimization problems that allow for indefinite Hessian approximations. Numerical experiments on a standard testing data set show that with a fixed computational time budget, the proposed methods achieve better results than the traditional limited-memory BFGS and the Hessian-free methods.

关键词： Alzheimer's disease   Hippocampal atrophy   Magnetic resonance imaging   Semi-automatic segmentation   Machine learning  

摘要： Alzheimer's disease is the most common form of dementia and is a serious health problem. The disease is expected to increase further in the upcoming years with the increase of the elderly population. Developing new treatments and diagnostic methods is getting more important. In this study, we focused on the early diagnosis of dementia in Alzheimer's disease via analysis of neuroimages. We analyzed the data diagnosed by the Alzheimer's Disease Neuroimaging Initiative (ADNI) protocol. The analyzed data were T1-weighted magnetic resonance images of 159 patients with Alzheimer's disease, 217 patients with mild cognitive impairment and 109 cognitively healthy older people. In this study, we propose that the volumetric reduction in the hippocampus is the most important indicator of Alzheimer's disease. There is not much research about the relationship between the volumetric reduction in the hippocampus and Alzheimer's disease. This volume information was calculated through semi-automatic segmentation software ITK-SNAP and a data set was created based on age, gender, diagnosis, and right and left hippocampal volume values. The diagnosis via hippocampal volume information was made by using machine learning techniques. By using this approach, we conclude that brain MRIs can be used to distinguish the patients with Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI) and Cognitive Normal (CN) from each other;while most of the studies were only able to distinguish AD from CN. Our results have revealed that our approach improves the performance of the computer-aided diagnosis of Alzheimer's disease.

关键词： 机器学习方法   层次聚类   因子分析方法   中医证候   四诊信息   处理与分析   量化标准   量化分析  

摘要： 传统的中医辨证诊疗主要基于"望、闻、问、切"得到的四诊信息,由于掺杂过多的医师主观因素,即使对同一个病人的辨证结果也可能不尽相同,因此如何建立一个科学而规范化的中医证候的量化标准是一个值得研究的课题。本文将机器学习中的层次聚类和因子分析方法应用于中医证候量化分析,通过对采集到的1499例典型高血压病例的处理与分析,实验结果表明,机器学习方法可以有效地挖掘中医证候中隐藏的信息,为中医辅助诊疗提供重要的途径。

关键词： Field capacity   Permanent wilting point   Deep learning   Artificial neural network   k-nearest neighbour  

摘要： This study evaluated the performance of deep learning (DL), artificial neural network (ANN) and k-nearest neighbour (kNN) models to estimate field capacity (FC) and permanent wilting point (PWP) using four combinations of soil data. The DL, ANN and kNN models are compared with the previous published pedotransfer functions (PTF). The data consist of 256 calcareous soil samples collected from Konya-Cumra plain, Turkey. The results demonstrated that the DL_a with inputs of soil texture components, bulk density, organic matter and lime contents, particle density and aggregate stability showed the best performances with coefficient of determination (R-2) of 0.829, correlation coefficient (r) of 0.911, mean absolute error (MAE) of 0.027 and relative root mean square error (RRMSE) 9.397 % in FC estimation for calcareous soil samples. For the PWP estimation of calcareous soil samples, the kNN_b with soil texture components, bulk density, organic matter and lime content and particle density indicated the best performance with the value of R-2 to 0.800, of r to 0.894, of MAE to 0.021 and RRMSE to 12.043 %. Lastly, the results showed that the DL, ANN and the kNN models perform better than the previously applied PTF for calcareous soils. Therefore, the DL model could be recommended for the estimation of FC when full soil data are available and the kNN model could be recommended for estimation of PWP with all combinations of soil data.