基于孪生卷积神经网络与三元组损失函数的图像识别模型

深度学习作为近年来迅速发展的全新领域,在科学研究与工业生产等方面受到了广泛的关注。其中,卷积神经网络(Convolutional neutral networks, CNN)作为深度学习中一种经典的神经网络架构,已在图像分类、人脸识别以及信号处理等领域得到了广泛的应用。在此基础上,本文对传统CNN结构进行改进,取消了CNN输出层用于普通分类的Softmax函数,采用基于孪生神经网络(Siamese neutral networks)的CNN架构,并使用三元组损失(Triplet Loss)作为图像分类的目标损失函数。为检验模型效果,我们在国际数据建模和数据分析竞赛平台Kaggle的座头鲸图像识别挑战赛上运用该模型。     

基于角度特征的分类网络

卷积神经网络(Convolutional Neural Networks,CNN)在图像分类任务中的卓越表现,使得其被广泛应用于计算机视觉的各个领域。图像分类模型精度与效率的提升,除了归功于网络结构的改变外,还有很大一部分原因来自于归一化技术以及分类损失函数的改进。在人脸识别任务中,随着精度的不断提升,分类损失函数从Softmax Loss到Triplet Loss,又从L-Softmax Loss到Arcface Loss,度量方式从几何度量发展到角度度量。度量方式的改变实际上是特征形式的变化,即特征形式从一般特征转变为角度特征。在Mnist数据集上,使用角度度量损失函数训练得到的特征点呈角度分布,同时准确率比几何度量高;将角度度量方式用更直接的角度特征来表示,训练得到的同类特征点呈直线分布,准确度也比一般角度度量更高。这不禁令人思考,在CNN分类模型中是否可以使用角度特征来代替一般特征。在CNN分类模型中,其主要架构往往由多个卷积层和一个或多个全连接层组成,通过统一卷积层与全连接层的归一化操作,得到角度卷积层与角度全连接层。在普通分类网络的基础上,用角度卷积层替换卷积层,用角度全连接层替换全连接层,可以得到一个由角度特征组成的角度分类网络。在Cifar-100数据集上,基于ResNet-32构造的角度分类网络相比原分类网络,分类准确率提高了2%,从而论证了角度特征在分类网络中的有效性。     

基于余弦距离损失与卷积神经网络的人脸识别

由于最近卷积神经网络在计算机视觉任务中的发展,深度人脸识别方法的性能得到了显著提高。现有的深度人脸模型将人脸识别任务视为一个分类任务或度量学习任务,旨在学习到具有区分度的人脸特征,但是很少能达到类内距离小、类间距离大的特点。损失函数作为监督信号,在卷积神经网络学习人脸特征时起到重要的作用。提出基于余弦距离的Softmax损失函数(Cosine Softmax Loss,CSL),让人脸特征区分度更高。使用相同的网络模型和训练数据集,在LFW和YTF等数据集上的实验结果验证了所提方法的优越性。     

改进残差神经网络实现心音分类

心音的变化可以较早表征心脏疾病体征,基于深度学习实现心音信号分类具有心脏疾病临床辅助无创诊断作用。秉承特征获取简易和深度学习模型简单理念,设计了融合CBAM注意力机制、Focal Loss损失函数和多尺度特征的ResNet152网络模型对PhysioNet/CinC 2016心音数据集进行深度学习;详细介绍了CBAM注意力机制以及在网络瓶颈结构中的融合方式、Focal Loss损失函数原理、多尺度特征获取方式,并设计了5组对比消融实验和横向对比实验。实验结果表明,CBAM注意力机制、Focal Loss损失函数和多尺度特征提高了ResNet152基线网络模型分类准确率。     

基于压缩采集与CNN的风电机轴承故障诊断

针对传统风电机轴承故障检测存在的采样数据量大、故障特征依赖主观选取的问题,提出了风电机轴承故障的信号压缩采集、自动提取特征及故障诊断的方法,解决了风电机轴承振动信号特征提取计算复杂、受先验知识影响较大的问题。首先基于梯度加速法(NAG)和QR分解理论对随机高斯观测矩阵进行优化,实现风电机轴承振动信号压缩采集;然后将压缩采集得到的数据作为卷积神经网络(CNN)的输入,利用卷积池化层提取压缩采集数据中的故障特征;最后,将得到的故障特征通过softmax分类器进行分类。仿真实验表明:该方法能够自动提取风电机轴承的故障特征,在保证较高故障诊断准确率的同时,缩短了网络训练时间。     

基于聚焦损失与残差网络的远程监督关系抽取

基于卷积神经网络的远程监督关系抽取方法提取的特征单一,且标准交叉熵损失函数未能较好处理数据集中正负样本比例不均衡的情况。为此,提出一种基于深度残差神经网络的远程监督关系抽取模型,通过改进交叉熵聚焦损失函数,提取句子中的深层语义特征,同时降低损失函数中负样本的权重,避免在NYT-Freebase标准数据集中引入NA关系类别的噪音。实验结果表明,该模型能增强深度残差神经网络对含噪音数据的表示学习能力,有效提高远程监督关系抽取任务的分类准确率。     

基于深度学习的带式输送机非煤异物识别方法

针对现有非煤异物图像识别法识别目标单一、模型缺乏定位能力等问题,提出一种基于深度学习的带式输送机非煤异物识别方法。该方法以目标检测算法YOLOv3为基础框架,采用Focal Loss函数替换YOLOv3模型中的交叉熵损失函数,对YOLOv3模型进行改进;通过调节最佳超参数(权重参数α和焦点参数γ)来平衡样本之间的比例,解决非煤异物样本不平衡问题,使模型在训练时更专注学习复杂目标样本特征,提高模型预测性能;搭建了异物数据集,并通过异物数据集对分类性能和速度进行实验。结果表明:Focal Loss函数在异物数据集中表现优于交叉熵损失函数,在γ=2,α=0.75时准确率提升5%,故最佳超参数为γ=2,α=0.75;改进后的YOLOv3模型对锚杆、角铁、螺母3种非煤异物的识别精确率分别提升了约4.7%,3.5%和6.8%,召回率分别提升了约6.6%,3.5%和6.0%;模型在2080Ti平台下每张图像预测类别与实际类别一致,且置信度在94%以上。     

面向三维模型视图特征提取的残差卷积网络优化

在已有残差卷积神经网络基础上,采用加权损失函数提高视图特征的可分性,提出面向三维模型视图特征提取的残差卷积网络优化算法.首先对三维模型进行多视图渲染得到二维视图;然后通过残差网络扩展模块加深网络深度;最后采用中心损失函数和交叉熵损失函数定义加权损失函数,解决交叉熵损失函数因为类内距离小于类间距离而导致的特征不可分问题.在ModelNet数据集上的实验结果表明,该算法提取到的特征在三维模型分类问题上性能表现优异.     

基于深度代价敏感CNN的年龄估计算法

针对年龄估计中样本数量不平衡及不同类间发生误分类时付出代价不同的问题,将代价敏感性嵌入深度学习框架中,提出基于深度代价敏感CNN的年龄估计算法.首先为每个年龄类别分别建立损失函数,解决训练样本的不平衡问题.然后,定义代价向量以反映不同类之间发生误分类而付出的代价差异性,构造逆交叉熵误差函数.最后,融合上述方法,为卷积神经网络(CNN)构造一个损失函数,使CNN在训练阶段学习针对年龄估计的鲁棒人脸表征.在不同种族的年龄估计标准图像集上的实验验证文中算法的有效性.     

基于最大相关熵准则的鲁棒度量学习算法

度量亦称距离函数,是度量空间中满足特定条件的特殊函数,一般用来反映数据间存在的一些重要距离关系.而距离对于各种分类聚类问题影响很大,因此度量学习对于这类机器学习问题有重要影响.受到现实存在的各种噪声影响,已有的各种度量学习算法在处理各种分类问题时,往往出现分类准确率较低以及分类准确率波动大的问题.针对该问题,本文提出一种基于最大相关熵准则的鲁棒度量学习算法.最大相关熵准则的核心在于高斯核函数,本文将其引入到度量学习中,通过构建以高斯核函数为核心的损失函数,利用梯度下降法进行优化,反复测试调整参数,最后得到输出的度量矩阵.通过这样的方法学习到的度量矩阵将有更好的鲁棒性,在处理受噪声影响的各种分类问题时,将有效地提高分类准确率.本文将在一些常用机器学习数据集（UCI）还有人脸数据集上进行验证实验.     

An improved convolutional neural network with an adaptable learning rate towards multi-signal fault diagnosis of hydraulic piston pump

Hydraulic piston pump is a vital component of hydraulic transmission system and plays a critical role in some modern industrials. On account of the deficiencies of traditional fault diagnosis in preprocessing of original data and feature extraction, the intelligent methods based on deep learning accomplish the automatic learning of fault information by integrating feature extraction and classification. As a popular deep learning model, convolutional neural network (CNN) has been demonstrated to be potent and effective in image classification. In this research, an improved intelligent method based on CNN with adapting learning rate is constructed for fault diagnosis of a hydraulic piston pump. Firstly, three raw signals are converted into two dimensional time–frequency images by continuous wavelet transform, including vibration signal, pressure signal and sound signal. Secondly, an improved deep CNN model is built with an adaptive learning rate strategy for identifying the different fault types. Moreover, t-distributed stochastic neighbor embedding is employed to visualize the distribution of features learned by the main layers of CNN model. Confusion matrix is used to analyze the classification accuracy of each fault type. Compared with the CNN model without adapting learning rate, the improved model achieves a higher accuracy based on the selected three kinds of signals. Experiments indicate that the improved CNN model can effectively and accurately identify various faults for a hydraulic piston pump.     

Intelligent fault diagnosis of rolling bearing using hierarchical convolutional network based health state classification

Rolling bearing tips are often the most susceptible to electro-mechanical system failure due to high-speed and complex working conditions, and recent studies on diagnosing bearing health using vibration data have developed an assortment of feature extraction and fault classification methods. Due to the strong non-linear and non-stationary characteristics, an effective and reliable deep learning method based on a convolutional neural network (CNN) is investigated in this paper making use of cognitive computing theory, which introduces the advantages of image recognition and visual perception to bearing fault diagnosis by simulating the cognition process of the cerebral cortex. The novel feature representation method for bearing data is first discussed using supervised deep learning with the goal of identifying more robust and salient feature representations to reduce information loss. Next, the deep hierarchical structure is trained in a robust manner that is established using a transmitting rule of greedy training layer by layer. Convolution computation, rectified linear units, and sub-sampling are applied for weight replication and reducing the number of parameters that need to be learned to improve the general feed-forward back propagation training. The CNN model could thus reduce learning computation requirements in the temporal dimension, and an invariance level of working condition fluctuation and ambient noise is provided by identifying the elementary features of bearings. A top classifier followed by a back propagation process is used for fault classification. Contrast experiments and analyses have been undertaken to delineate the effectiveness of the CNN model for fault classification of rolling bearings.     

One-dimensional convolutional auto-encoder-based feature learning for fault diagnosis of multivariate processes

Noise and high-dimension of process signals decrease effectiveness of those regular fault detection and diagnosis models in multivariate processes. Deep learning technique shows very excellent performance in high-level feature learning from image and visual data. However, the large labeled data are required for deep neural networks (DNNs) with supervised learning like convolutional neural network (CNN), which increases the time cost of model construction significantly. A new DNN model, one-dimensional convolutional auto-encoder (1D-CAE) is proposed for fault detection and diagnosis of multivariate processes in this paper. 1D-CAE is utilized to learn hierarchical feature representations through noise reduction of high-dimensional process signals. Auto-encoder integrated with convolutional kernels and pooling units allows feature extraction to be particularly effective, which is of great importance for fault detection and diagnosis in multivariate processes. The comparison between 1D-CAE and other typical DNNs illustrates effectiveness of 1D-CAE for fault detection and diagnosis on Tennessee Eastman Process and Fed-batch fermentation penicillin process. The proposed method provides an effective platform for deep-learning-based process fault detection and diagnosis of multivariate processes.     

Automatic determination of digital modulation types with different noises using Convolutional Neural Network based on time–frequency information

In this study, a novel digital modulation classification model has been proposed for automatically recognizing six different modulation types including amplitude shift keying (ASK), frequency shift keying (FSK), phase-shift keying (PSK), quadrate amplitude shift keying (QASK), quadrate frequency shift keying (QFSK), and quadrate phase-shift keying (QPSK). The determination of modulation type is significant in military communication, satellite communication systems, and submarine communication. To classify the modulation types, we have proposed a two-stage hybrid method combining short-time Fourier transform (STFT) and convolutional neural network (CNN). In the first stage, as the data source, the time–frequency information from these modulation signals have been extracted with STFT. This information has been obtained as 2D images to feed the input of the CNN deep learning method. In the second stage, the obtained 2D time–frequency information has been given to the input of the CNN algorithm to classify the modulation types. In this work, noises at various SNR values from 0 dB to 25 dB were created and added to the modulated signals. Even in the presence of noise, the proposed hybrid deep learning model achieved excellent results in the noised-modulation signals.     

基于离散小波变换和随机森林的轴承故障诊断研究

针对不同工况下数据特征选择困难和单一分类器在滚动轴承故障诊断中识别率较低等问题,提出了一种基于离散小波变换和随机森林相结合的滚动轴承故障诊断方法。该方法首先利用离散小波变换分解振动信号,得到n层近似系数;然后创新性地采用sigmoid熵构造出n维特征向量,sigmoid熵能较好地提取非平稳信号的特征,提高诊断准确率;最后采用随机森林对滚动轴承不同故障信号进行分类。实验采用西储凯斯大学轴承数据中心网站提供的轴承数据,与传统分类器(KNN和SVM)以及单个分类回归树CART进行对比分析,结果表明该方法具有更好的诊断效果。     

Acoustic Emission Recognition Based on a Three-Streams Neural Network with Attention

Acoustic emission (AE) is a nondestructive real-time monitoring technology, which has been proven to be a valid way of monitoring dynamic damage to materials. The classification and recognition methods of the AE signals of the rotor are mostly focused on machine learning. Considering that the huge success of deep learning technologies, where the Recurrent Neural Network (RNN) has been widely applied to sequential classification tasks and Convolutional Neural Network (CNN) has been widely applied to image recognition tasks. A novel three-streams neural network (TSANN) model is proposed in this paper to deal with fault detection tasks. Based on residual connection and attention mechanism, each stream of the model is able to learn the most informative representation from Mel Frequency Cepstrum Coefficient (MFCC), Tempogram, and short-time Fourier transform (STFT) spectral respectively. Experimental results show that, in comparison with traditional classification methods and single-stream CNN networks, TSANN achieves the best overall performance and the classification error rate is reduced by up to 50%, which demonstrates the availability of the model proposed.     

基于空洞卷积神经网络的旋转机械故障诊断方法

由于旋转机械的振动信号具有非平稳、复杂多样、数据量大的特点，传统的方法难以较好地实现旋转机械故障诊断。近年来，基于深度学习的故障诊断算法发展迅速，其中，卷积神经网络(Convolutional Neural Network,CNN)由于可实现自动提取特征、运算效率高等优点受到广泛关注，但在识别准确率等方面仍然存在部分问题。为实现多传感器监测状态下的旋转机械故障诊断，在经典卷积神经网络结构的基础上，引入了多通道数据融合处理、空洞卷积层、批标准化处理、PReLU激活函数、全局平均池化层等改进方法，构造了一种新型的、高效的空洞卷积神经网络（Atrous Convolution-Convolutional Neural Network,AC-CNN），并基于该模型进行了旋转机械故障诊断实验。实验结果表明，提出的故障诊断模型分类准确率可达99%以上，对比其他神经网络方法具有明显优势。     

模糊熵和深度学习在精神分裂症中的应用研究

精神分裂症是一种常见的重性精神疾病,多年来严重影响人类的生活质量,因此,对该病的准确诊断是治疗疾病的前提。针对以上问题,提出一种基于大脑复杂性和深度学习的精神分裂症脑电信号（EEG）分类方法,旨在发现隐藏在数据中的分布式特征。与忽略空间信息的标准脑电数据分析技术相反,首先将脑电信号的时间序列进行分频处理,并将每个频段的时间序列用模糊熵（FuzzyEn）进行特征提取,按照电极的空间位置构成特征向量,并将特征向量输入到卷积神经网络（CNN）中训练分类模型,自动识别受试者是否患病。实验结果表明,基于模糊熵和卷积神经网络的分类方法是有效的,分类准确率达到了99.16%。     

基于卷积神经网络的滚动轴承故障诊断方法

为了简单、准确地进行轴承故障诊断,结合深度学习理论,对基于卷积神经网络的滚动轴承故障诊断方法进行了研究。首先,选用了结构相对简单的LeNet5卷积神经网络;然后,对轴承振动信号原始数据进行截取和归一化处理后直接生成生成二维矩阵作为神经网络输入;接着,优选卷积核大小、批大小、学习率及迭代次数等网络模型参数;最后,应用sigmoid函数进行多标签分类。实验结果表明,该方法能有效识别正常状态及不同损伤程度下的内圈、外圈、滚动体故障状态,识别准确率达到99.50%以上水平。基于卷积神经网络的滚动轴承故障诊断方法不仅在一定程度上可以简化故障诊断的过程,而且可以充分利用卷积神经网络模型的优势实现高效准确地故障诊断。     

基于卷积神经网络和代价敏感的不平衡图像分类方法

针对不平衡图像分类中少数类查全率低、分类结果总代价高,以及人工提取特征主观性强而且费时费力的问题,提出了一种基于Triplet-sampling的卷积神经网络（Triplet-sampling CNN）和代价敏感支持向量机（CSSVM）的不平衡图像分类方法——Triplet-CSSVM。该方法将分类过程分为特征学习和代价敏感分类两部分。首先,利用误差公式为三元损失函数的卷积神经网络端对端地学习将图像映射到欧几里得空间的编码方法;然后,结合采样方法重构数据集,使其分布平衡化;最后,使用CSSVM分类算法给不同类别赋以不同的代价因子,获得最佳代价最小的分类结果。在深度学习框架Caffe上使用人像数据集FaceScrub进行实验。实验结果表明,所提方法在1∶3的不平衡率下,与VGGNet-SVM方法相比,少数类的精确率提高了31个百分点,召回率提高了71个百分点。