YOLSO: You Only Look Small Object

Small object detection is challenging and far from satisfactory. Most general object detectors suffer from two critical issues with small objects: (1) Feature extractor based on classification network cannot express the characteristics of small objects reasonably due to insufficient appearance information of targets and a large amount of background interference around them. (2) The detector requires a much higher location accuracy for small objects than for general objects. This paper proposes an effective and efficient small object detector YOLSO to address the above problems. For feature representation, we analyze the drawbacks in previous backbones and present a Half-Space Shortcut(HSSC) module to build a background-aware backbone. Furthermore, a coarse-to-fine Feature Pyramid Enhancement(FPE) module is introduced for layer-wise aggregation at a granular level to enhance the semantic discriminability. For loss function, we propose an exponential L1 loss to promote the convergence of regression, and a focal IOU loss to focus on prime samples with high classification confidence and high IOU. Both of them significantly improves the location accuracy of small objects. The proposed YOLSO sets state-of-the-art results on two typical small object datasets, MOCOD and VeDAI, at a speed of over 200 FPS. In the meantime, it also outperforms the baseline YOLOv3 by a wide margin on the common COCO dataset.     

Robust optimal operation of continuous catalytic reforming process under feedstock uncertainty

The continuous catalytic regenerative (CCR) reforming process is one of the most significant sources of hydrogen production in the petroleum refining process. However, the fluctuations in feedstock composition and flow rate could significantly affect both product distribution and energy consumption. In this study, a robust deviation criterion based multi-objective optimization approach is proposed to perform the optimal operation of CCR reformer under feedstock uncertainty, with simultaneous maximization of product yields and minimization of energy consumption. Minimax approach is adopted to handle these uncertain objectives, and the Latin hypercube sampling method is then used to calculate these robust deviation criteria. Multi-objective surrogate-based optimization methods are next introduced to effectively solve the robust operational problem with high computational cost. The level diagram method is finally utilized to assist in multi-criteria decision-making. Two robust operational optimization problems with different objectives are solved to demonstrate the effectiveness of the proposed method for robust optimal operation of the CCR reforming process under feedstock uncertainty.     

Machine learning algorithm selection for windage alteration fault diagnosis of mine ventilation system

Machine learning algorithms have been widely used in mine fault diagnosis. The correct selection of the suitable algorithms is the key factor that affects the fault diagnosis. However, the impact of machine learning algorithms on the prediction performance of mine fault diagnosis models has not been fully evaluated. In this study, the windage alteration faults (WAFs) diagnosis models, which are based on K-nearest neighbor algorithm (KNN), multi-layer perceptron (MLP), support vector machine (SVM), and decision tree (DT), are constructed. Furthermore, the applicability of these four algorithms in the WAFs diagnosis is explored by a T-type ventilation network simulation experiment and the field empirical application research of Jinchuan No. 2 mine. The accuracy of the fault location diagnosis for the four models in both networks was 100%. In the simulation experiment, the mean absolute percentage error (MAPE) between the predicted values and the real values of the fault volume of the four models was 0.59%, 97.26%, 123.61%, and 8.78%, respectively. The MAPE for the field empirical application was 3.94%, 52.40%, 25.25%, and 7.15%, respectively. The results of the comprehensive evaluation of the fault location and fault volume diagnosis tests showed that the KNN model is the most suitable algorithm for the WAFs diagnosis, whereas the prediction performance of the DT model was the second-best. This study realizes the intelligent diagnosis of WAFs, and provides technical support for the realization of intelligent ventilation.     

Tabular data: Deep learning is not all you need

A key element in solving real-life data science problems is selecting the types of models to use. Tree ensemble models (such as XGBoost) are usually recommended for classification and regression problems with tabular data. However, several deep learning models for tabular data have recently been proposed, claiming to outperform XGBoost for some use cases. This paper explores whether these deep models should be a recommended option for tabular data by rigorously comparing the new deep models to XGBoost on various datasets. In addition to systematically comparing their performance, we consider the tuning and computation they require. Our study shows that XGBoost outperforms these deep models across the datasets, including the datasets used in the papers that proposed the deep models. We also demonstrate that XGBoost requires much less tuning. On the positive side, we show that an ensemble of deep models and XGBoost performs better on these datasets than XGBoost alone.     

基于多通道的ICU脑血管疾病死亡风险预测模型

死亡风险预测指根据病人临床体征监测数据来预测未来一段时间的死亡风险。对于ICU病患，通过死亡风险预测可以有针对性地对病人做出临床诊断，以及合理安排有限的医疗资源。基于临床使用的MEWS和Glasgow昏迷评分量表，针对ICU病人临床监测的17项生理参数，提出一种基于多通道的ICU脑血管疾病死亡风险预测模型。引入多通道概念应用于BiLSTM模型，用于突出每个生理参数对死亡风险预测的作用。采用Attention机制用于提高模型预测精度。实验数据来自MIMIC [Ⅲ]数据库，从中提取3?080位脑血管疾病患者的16?260条记录用于此次研究，除了六组超参数实验之外，将所提模型与LSTM、Multichannel-BiLSTM、逻辑回归（logistic regression）和支持向量机（support vector machine, SVM）四种模型进行了对比分析，准确率Accuracy、灵敏度Sensitive、特异性Specificity、AUC-ROC和AUC-PRC作为评价指标，实验结果表明，所提模型性能优于其他模型，AUC值达到94.3%。     

基于VMD和BSA-KELM的高陡边坡位移预测模型研究

边坡位移的时间序列曲线存在复杂的非线性特性,传统的预测模型精度不足以满足预测要求。为此提出了基于变分模态分解的鸟群优化-核极限学习机的预测模型,并用于河北省某水泥厂的边坡位移预测。该方法首先采用VMD把边坡位移序列分解为一系列的有限带宽的子序列,再对各子序列分别采用相空间重构并用核极限学习机预测,采用鸟群算法优化相空间重构的嵌入维度和KELM中惩罚系数和核参数三个数值,以取得最优预测模型。最后将各个子序列预测值叠加,得到边坡位移的最终预测值。结果表明：和KELM、BSA-KELM、EEMD-BSA-KELM模型相比,基于VMD的BSA-KELM预测精度更高,为边坡位移的预测提供一种有效的方法。     

Development of high pressure membraneless alkaline electrolyzer

A technology for cyclic generation of hydrogen and oxygen using electrodes made of variable valency material that does not need the use of separating ion-exchange membranes is presented. The technological solution enables to fabricate electrolyzers for uninterrupted producing high-pressure hydrogen with reduced energy intensity of the production. The total work for compressing 1 m³ of hydrogen and 0.5 m³ of oxygen has been estimated. Results of investigation of influence of discrete supply of DC current to the electrolysis cell, in order to improve the processes of gas evolution and to simplify the power systems of the electrolysis plant, have been considered. There is also considered an electrolysis installation equipped with a thermosorption compressor in which LaNi₅ is used as a hydride-forming compound. The comparative characteristics of the developed electrolyzer and the currently used hydrogen generators are given.     

Designing optical glasses by machine learning coupled with a genetic algorithm

Engineering new glass compositions have experienced a sturdy tendency to move forward from (educated) trial-and-error to data- and simulation-driven strategies. In this work, we developed a computer program that combines data-driven predictive models (in this case, neural networks) with a genetic algorithm to design glass compositions with desired combinations of properties. First, we induced predictive models for the glass transition temperature (T_g) using a dataset of 45,302 compositions with 39 different chemical elements, and for the refractive index (n_d) using a dataset of 41,225 compositions with 38 different chemical elements. Then, we searched for relevant glass compositions using a genetic algorithm informed by a design trend of glasses having high n_d (1.7 or more) and low T_g (500 °C or less). Two candidate compositions suggested by the combined algorithms were selected and produced in the laboratory. These compositions are significantly different from those in the datasets used to induce the predictive models, showing that the used method is indeed capable of exploration. Both glasses met the constraints of the work, which supports the proposed framework. Therefore, this new tool can be immediately used for accelerating the design of new glasses. These results are a stepping stone in the pathway of machine learning-guided design of novel glasses.     

Comparative analysis of inbreeding parameters and runs of homozygosity islands in 2 Italian autochthonous cattle breeds mainly raised in the Parmigiano-Reggiano cheese production region

Reggiana and Modenese are autochthonous cattle breeds, reared in the North of Italy, that can be mainly distinguished for their standard coat color (Reggiana is red, whereas Modenese is white with some pale gray shades). Almost all milk produced by these breeds is transformed into 2 mono-breed branded Parmigiano-Reggiano cheeses, from which farmers receive the economic incomes needed for the sustainable conservation of these animal genetic resources. After the setting up of their herd books in 1960s, these breeds experienced a strong reduction in the population size that was subsequently reverted starting in the 1990s (Reggiana) or more recently (Modenese) reaching at present a total of about 2,800 and 500 registered cows, respectively. Due to the small population size of these breeds, inbreeding is a very important cause of concern for their conservation programs. Inbreeding is traditionally estimated using pedigree data, which are summarized in an inbreeding coefficient calculated at the individual level (F_PED). However, incompleteness of pedigree information and registration errors can affect the effectiveness of conservation strategies. High-throughput SNP genotyping platforms allow investigation of inbreeding using genome information that can overcome the limits of pedigree data. Several approaches have been proposed to estimate genomic inbreeding, with the use of runs of homozygosity (ROH) considered to be the more appropriate. In this study, several pedigree and genomic inbreeding parameters, calculated using the whole herd book populations or considering genotyping information (GeneSeek GGP Bovine 150K) from 1,684 Reggiana cattle and 323 Modenese cattle, were compared. Average inbreeding values per year were used to calculate effective population size. Reggiana breed had generally lower genomic inbreeding values than Modenese breed. The low correlation between pedigree-based and genomic-based parameters (ranging from 0.187 to 0.195 and 0.319 to 0.323 in the Reggiana and Modenese breeds, respectively) reflected the common problems of local populations in which pedigree records are not complete. The high proportion of short ROH over the total number of ROH indicates no major recent inbreeding events in both breeds. ROH islands spread over the genome of the 2 breeds (15 in Reggiana and 14 in Modenese) identified several signatures of selection. Some of these included genes affecting milk production traits, stature, body conformation traits (with a main ROH island in both breeds on BTA6 containing the ABCG2, NCAPG, and LCORL genes) and coat color (on BTA13 in Modenese containing the ASIP gene). In conclusion, this work provides an extensive comparative analysis of pedigree and genomic inbreeding parameters and relevant genomic information that will be useful in the conservation strategies of these 2 iconic local cattle breeds.     

Feature selection methods for root-cause analysis among top-level product attributes

Manufacturing companies not only strive to deliver flawless products but also monitor product failures in the field to identify potential quality issues. When product failures occur, quality engineers must identify the root cause to improve any affected product and process. This root-cause analysis can be supported by feature selection methods that identify relevant product attributes, such as manufacturing dates with an increased number of product failures. In this paper, we present different methods for feature selection and evaluate their ability to identify relevant product attributes in a root-cause analysis. First, we compile a list of feature selection methods. Then, we summarize the properties of product attributes in warranty case data and discuss these properties regarding the challenges they pose for machine learning algorithms. Next, we simulate datasets of warranty cases, which emulate these product properties. Finally, we compare the feature selection methods based on these simulated datasets. In the end, the univariate filter information gain is determined to be a suitable method for a wide range of applications. The comparison based on simulated data provides a more general result than other publications, which only focus on a single use case. Due to the generic nature of the simulated datasets, the results can be applied to various root-cause analysis processes in different quality management applications and provide a guideline for readers who wish to explore machine learning methods for their analysis of quality data.