基于空洞卷积神经网络与注意力机制GRU的滚动轴承故障诊断

针对传统故障诊断方法需要人工提取特征的不足,以及大数据下滚动轴承故障振动信号自适应特征提取与智能诊断问题,利用空洞卷积神经网络(DCNN)可以在不增加计算量的基础上兼顾不同尺度空间特征的能力、门控循环单元(GRU)善于从动态变化的序列数据中学习到时间上的关联性的能力,提出了一种将DCNN、注意力机制和GRU多路径融合的端到端故障诊断方法。首先利用DCNN从原始数据中自动提取时序信号特征,然后将注意力机制(Attention)的GRU通路和DCNN通路进行融合,最后将提取到的特征融合之后送入分类层进行分类。试验结果表明,所提方法的诊断准确率平均为98.75%,高于比较方法,更加适用于滚动轴承故障诊断。     

基于一维卷积神经网络的儿童睡眠分期

高质量睡眠与儿童的身体发育、认知功能、学习和注意力密切相关,由于儿童睡眠障碍的早期症状不明显,需要进行长期监测,因此急需找到一种适用于儿童睡眠监测,且能够提前预防和诊断此类疾病的方法。多导睡眠图(Polysomnography,PSG)是临床指南推荐的睡眠障碍基本检测方法,通过观察PSG各睡眠期间的变化和规律,对睡眠质量评估和睡眠障碍识别具有基础作用。本文对儿童睡眠分期进行了研究,利用多导睡眠图记录的单通道脑电信号,在Alexnet的基础上,用一维卷积代替二维卷积,提出一种1D-CNN结构,由5个卷积层、3个池化层和3个全连接层组成,并在1D-CNN中添加了批量归一化层（Batch normalization layer）,保持卷积核的大小保持不变。针对数据集少的情况,采用了重叠的方法对数据集进行了扩充。实验结果表明,该模型儿童睡眠分期的准确率为84.3%。通过北京市儿童医院的PSG数据获得的归一化混淆矩阵,可以看出,Wake、N2、N3和REM期睡眠的分类性能很好。对于N1期睡眠,存在将N1期睡眠被误分类为Wake、N2和REM期睡眠的情况,因此以后的工作应重点提升N1期睡眠的准确性。总体而言,对于基于带有睡眠阶段标记的单通道EEG的自动睡眠分期,本文提出的1D-CNN模型可以实现针对于儿童的自动睡眠分期。在未来的工作中,仍需要研究开发更适合于儿童的睡眠分期策略,在更大数据量的基础上进行实验。      

基于任务态和静息态功能核磁共振信号的抑郁症识别

为了提高抑郁症识别的准确率,将功能核磁共振成像的任务态数据和静息态数据相结合,建立基于数据驱动的模型以提取识别特征.在没有任何先验知识的条件下,采用独立成分分析法提取任务态数据和静息态数据的独立成分;然后,利用相关遍历分析法获取功能信号集,利用频谱分析法识别并获取功能信号成分;最后,将功能信号成分作为贝叶斯分类器的特征输入,完成分类.结果表明,利用该方法提取出的功能信号成分能很好地将抑郁症患者和健康者区分开,整体识别准确率达到77.27%,抑郁症患者识别准确率达到83.33%,健康者识别准确率达到70.00%.实验结果证明了这一方法的有效性及优越性.     

基于自训练的多标签岩矿石薄片分类方法

岩矿石薄片识别是一项专业性要求极高的任务,人工识别常出现不可避免的主观错误,且效率极低。深度学习图像识别技术是可以高效进行岩矿石薄片识别的方法,但训练深度学习模型需要大量标注数据,因此如何高效利用有限标注数据具有重要意义。通过采用多标签分类方法,在有标签数据集上先训练一个分类器,然后使用该分类器为大量无标注的岩矿石薄片生成伪标签,最后使用有标签的训练数据和所有无标签数据重新训练模型。结果表明,采用多标签分类方法识别岩矿石薄片结构及矿物是可行的,同时使用半监督学习方法训练模型,在不进行大量人工标注的情况下,可提高该模型的泛化能力。     

基于改进降噪自编码器半监督学习模型的热轧带钢水梁印识别算法

针对水梁印识别困难且工作量大问题,提出一种基于改进降噪自编码器半监督学习模型的热轧带钢水梁印识别算法。该算法在降噪自编码器（Denoising auto-encoder, DAE）的基础上对编码层的每一层添加随机噪声,在隐藏层后添加分类层,并对数据添加伪标签,在解码的同时进行分类训练,使得DAE具有半监督学习能力。通过提取热轧带钢粗轧出口温度数据中的温差特征,用相应特征对模型进行训练。实验结果表明,算法能够准确识别出带钢的水梁印,在模型精确度上,与主流分类识别模型对比,提出的模型在带标签样本数量较小时,分类精度相比其他模型高5.0%～10.0%;在带标签样本数量较大时,提出的模型分类精度达到93.8%,现场能够根据模型的识别结果提高生产效率。      

基于一维卷积特征与手工特征融合的集成超限学习机心跳分类方法

融合手工特征和深度特征,提出了一种集成超限学习机心跳分类方法。手工提取的特征明确地表征了心电信号的特定特性,如相邻心跳时间间隔反映了心跳信号的时域特性,小波系数反映了心跳信号的时频特性。同时设计了一维卷积神经网络对心跳信号特征进行自动提取。基于超限学习机（Extreme leaning machine,ELM）,将上述特征融合进行心跳分类。由于ELM初始参数的随机给定可能导致其性能不稳定,进一步提出了一种基于袋装（Bagging）策略的多个ELM集成方法,使分类结果更加稳定且模型泛化能力更强。利用麻省理工心律失常公开数据集对所提方法进行了验证,分类准确率达到了99.02%,实验结果也表明基于融合特征的分类准确率高于基于单独特征的分类准确率。      

基于SE注意力机制的废钢分类评级方法

为了解决传统人工方法对废钢分类评级人为因素干扰大且效率低下等问题,提出基于挤压?激励（Squeeze?Excitation,SE）注意力机制构建废钢分类评级的深度学习网络模型,并对采集到的废钢卸载过程图像进行模型训练和验证。首先,搭建物理尺寸比例为1∶3废钢质量查验物理模型,采用高分辨率视觉传感器模拟采集货车卸载废钢作业场景下不同废钢的形貌特征;然后,对采集到的废钢图像使用跨阶段局部网络进行特征提取,利用空间金字塔结构解决特征丢失问题,采用注意力机制关注通道间的相关性;最后,在包含7个标签分类的两个数据集进行模型训练与验证。实验表明:该模型能够有效地对不同级别的废钢进行自动评级判定,全类别准确率达到83.7%,全类别平均精度为88.8%,在准确性方面相比于传统人工验质方法具有显著优势,解决了废钢入库过程中质量评价的公正性难题。      

基于BiLSTM-Attention的钢板表面手写板号识别算法

国内钢铁企业生产厂的信息化物料跟踪大都依赖于钢板号。由于生产流程复杂,急需高准确率的板号在线识别技术。自然场景下机器喷号的识别技术较成熟,但复杂场景下的手写板号难以实现自动识别。针对复杂工作场景下钢板表面手写板号特点,提出一种以BiLSTM-Attention为主体结构的深度学习算法。首先结合复杂场景,对图像数据进行预处理,保证模型输入图片质量;然后利用残差神经网络(ResNet)提取图片特征、利用双向长短期记忆网络(BiLSTM)提取基于图像的序列特征;最后基于注意力机制捕获序列内的信息流,对每个字符的特征进行整合,形成文本特征向量以预测输出序列。经现场测试,实现钢板表面手写板号识别任务准确率达86.15%,结果表明算法可行有效,满足实际生产需求。     

储集层测井信号的非线性混沌特性

对测井信号的特征提取采用非线性分析方法.利用非平稳信号分析方法--游程检验法,分析松辽盆地某井储集层测井数据的非平稳特性;在此基础上,采用非线性混沌方法对测井数据进行最大Lyapunov指数提取.结果表明:储集层测井序列具有非平稳特性;最大的Lyapunov指数大于零,表明了储集层测井序列具有混沌特征.为此,储集层测井信息的提取可采用非线性混沌理论方法,以便更有效地描述测井信息特征.     

基于红外图像的矿石传送带托辊异常检测

为了解决传统的传送带托辊异常检测方法效率低、实时性差等问题,提出一种基于红外图像识别的托辊异常检测模型。通过现场采集并使用标签平滑和Mosaic数据增强处理对托辊红外图像数据集进行扩充,降低模型的训练成本。在特征提取模块提出使用GhostNet骨干特征提取网络,能够有效地降低特征提取所需成本。在特征融合模块,提出使用SPP-Net模块优化PaNet特征融合网络,增加模型的感受野。通过深度可分离卷积块简化模型结构,降低模型的计算量和参数量,并通过LeakyReLU激活函数提高模型的学习能力。试验结果表明：该检测模型能够有效识别托辊异常。在实际检测中,该方法在托辊检测中平均准确率达到94.9%,检测速度达到39.2 FPS,为矿山传送带托辊的准确高效巡检提供了保障。     

Data-Driven Approach Using Supervised Learning for Predicting Endpoint Temperature of Molten Steel in the Electric Arc Furnace

This study presents the development of a decision support system that focuses on predicting the endpoint temperature of molten steel to manage the process of an electric arc furnace more systematically. The decision support system leverages a data-driven approach that consists of the following modules: 1) a data preprocessing module that specifically includes raw data filtering, feature engineering, and outlier detection; 2) a feature selection module based on domain knowledge; 3) regression modeling module that employs a supervised learning algorithm to forecast an endpoint temperature; and 4) sensitivity analysis module to identify the correlation between input and output metric. The applicability of the system is demonstrated through a validation study using real-world operational data from Hyundai Steel located in Pohang, South Korea. The validation results show that the endpoint temperatures predicted by the system are evenly scattered to a perfect-fit line within 5% errors of the actual temperatures. The results also indicate that CaO, power, and melting score have the most significant impact on the endpoint temperature, in which temperature decreases as CaO increases and increases as the power and melting score increase.     

Stochastic complexity measures for physiological signal analysis

Traditional feature extraction methods describe signals in terms of amplitude and frequency. This paper takes a paradigm shift and investigates four stochastic-complexity features. Their advantages are demonstrated on synthetic and physiological signals; the latter recorded during periods of Cheyne-Stokes respiration, anesthesia, sleep, and motor-cortex investigation.     

基于多域流形的行星齿轮箱局部故障识别

行星齿轮箱振动信号包含多种频率成分和噪声干扰，频谱具有复杂的边带结构，容易对故障识别造成误导甚至引起错判.在不同故障状态下，行星齿轮箱振动信号的多域特征量将偏离正常范围且偏离程度不同，根据这一特点，提取振动信号的时域、频域特征参量用于故障识别.为了避免传统分析方法中负频率及虚假模态问题，增强对噪声干扰的鲁棒性，采用局部均值分解法将信号自适应地分解为单分量之和，提取时频域单分量瞬时幅值能量.针对多域特征空间构造过程中出现的高维及非线性问题，采用流形学习对数据进行降维处理.提出基于改进的虚假近邻点的本征维数估计及最优k邻域确定方法，并通过等距映射对多域特征空间进行降维分析.对于行星齿轮箱实验信号，根据样本流形特征聚类结果，分别识别出了太阳轮、行星轮和齿圈的局部故障，从而验证了上述方法的有效性.      

基于时-频分析的步态模式自动分类

针对不同路况和运动模式下的高维、非线性、强耦合和高时变下肢加速度信号的识别问题,提出了一种基于时——频分析的步态模式自动分类方案.利用三轴加速度传感器采集运动时小腿在矢状面、冠状面和横切面的加速度信号,利用五阶Daubechies小波基对其进行特征提取,并采用线性判别式分析进行降维,最后利用决策树和支持向量机对得到的精简步态特征进行模式分类.实验结果显示两种分类器的总体分类准确率均达到90%以上,个别步态分类可达到100%,验证了特征提取和降维方法的合理性和有效性.      

The costs of supervised classification: The effect of learning task on conceptual flexibility.

Research has shown that learning a concept via standard supervised classification leads to a focus on diagnostic features, whereas learning by inferring missing features promotes the acquisition of within-category information. Accordingly, we predicted that classification learning would produce a deficit in people's ability to draw novel contrasts—distinctions that were not part of training—compared with feature inference learning. Two experiments confirmed that classification learners were at a disadvantage at making novel distinctions. Eye movement data indicated that this conceptual inflexibility was due to (a) a narrower attention profile that reduces the encoding of many category features and (b) learned inattention that inhibits the reallocation of attention to newly relevant information. Implications of these costs of supervised classification learning for views of conceptual structure are discussed. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Acoustic hallucinations as a symptom of a REM sleep-associated parasomnia

This 52-year-old man suffered from auditory hallucinations that occurred during brief episodes of sleep paralysis at the end of REM sleep periods. During these episodes the patient experienced a dissociated state of consciousness with REM sleep intrusions into wakefulness. The occurrence of this mixed state, and of excessive sleep-onset REM periods during daytime polysomnography (MSLT = Multiple Sleep Latency Test), point to a disorder of REM sleep generation. The existence of narcolepsy could be ruled out. The observation of REM sleep-associated hallucinations has been reported earlier. In the presented polysomnographic sleep studies the existence of a REM sleep associated parasomnia characterised by hallucinations and sleep paralysis could be confirmed.     

Polysomnographic and subjective sleep predictors of alcoholic relapse

Previous studies indicate that subjectively reported and objectively measured sleep abnormalities at baseline can increase the risk of relapse in treated alcoholics. However, previous studies did not include both subjective and objective sleep measures in the same group of patients. We utilized polysomnography and the Sleep Disorders Questionnaire to determine if baseline polysomnography increased the ability to predict relapse beyond the prediction with subjective measures alone, after controlling for nonsleep variables that were associated with relapse. We followed 74 patients with a DSM-III-R diagnosis of alcohol dependence, of whom 36 relapsed to at least some drinking during an average follow-up interval of 5 months. Univariate analyses revealed that relapsed patients did not differ from abstinent patients at baseline in demographics or psychiatric co-morbidity, but they had more prior treatment episodes for alcoholism, more difficulty falling asleep, more complaints of abnormal sleep, and, on polysomnography, longer sleep latencies, shorter rapid eye movement sleep latencies, and less stage 4 sleep percentage than abstinent patients. With a series of logistic regression analyses, which controlled for age and gender, we demonstrated that sleep measures improved the prediction model compared with nonsleep variables alone, and that polysomnography-measured sleep latency was the most significant predictor variable. We conclude that subjective and objective measures of baseline sleep are predictors of relapse in treated alcoholic patients. These data also suggest that neurophysiological dysfunction contributes strongly to the etiology of relapse. Finally, sleep disturbance warrants clinical attention as a target of alcoholism treatment.     

Clinical aspects of sleep disorders--experiences with 817 patients of an ambulatory sleep clinic; comment

Between 1992, the year in which the Sleep Out-Patient Clinic at the Department of Psychiatry, University of Vienna, Allgemeines Krankenhaus (General Hospital) Vienna, was established, and 1996, 817 patients (58% females, average age 52 years; 42% males, average age 48 years) were treated for sleep disorder. According to the International Statistical Classification of Diseases and Related Health Problems (ICD-10) of the World Health Organization (WHO), 70% of the patients presented with a non-organic sleep disorder and 30% with an organic sleep disorder as main diagnosis. Non-organic insomnia was by far the most frequently diagnosed sleep disorder (48%), while within the organic sleep disorders sleep apnea was dominant (12%). In regard to the additional non-organic (mental disorder) diagnoses rounding off the clinical picture, neurotic, stress related, and somatoform disorders were the most common (41%), followed by affective disorders (31%) and mental and behavioural disorders due to intake of psychoactive substances, e.g. alcohol, drugs (15%). Additional organic diagnoses related to sleep disorders involved primarily endocrine disorders such as adipositas (23%), followed by cardiovascular disorders (19%), and primary snoring (17%). The sleep out-patient clinic has at its disposal a supportive diagnostic armamentarium such as all-night sleep polysomnography, 24-hour polysomnography, the Multiple Sleep Latency Test, EEG and EEG-mapping in the affiliated sleep laboratory, the evaluation of event-related potentials (P300) and actometry in the psychophysiological laboratory, as well as psychological and psychophysiological tests in the clinical psychodiagnostic laboratory, in order to determine the right treatment or preventive measures for the individual patients.     

基于小波包分析的AOD炉喷溅预报特征信息提取

研究了用小波包分析方法从炉口音频信号中提取AOD炉喷溅预报特征信息的方法。采用db10（小波基函数）小波对喷溅发生前的特征信号进行4层小波包分解,结合快速傅里叶变换法及小波尺度谱进行时频特征分析,并研究了其各频带分解信号的能量比例特点。结果表明,喷溅前40s信号的主频值较正常信号有明显降低,0~312Hz与312~625Hz频段信号能量值比例变化显著。而且低频重构信号可以极好地滤除多种现场干扰,说明该时频特征可以作为准确预报喷溅的特征向量。最后,通过实验确定了8个特征向量值并分别与喷溅或正常信号的特征向量进行相关性比较,验证得出相关度0.95可作为喷溅预报的判定阈值。从而实现了喷溅预报特征信号的准确提取并可转化为计算机容易识别的数值特征。