基于NVDLA与FPGA结合的神经网络加速器平台设计

随着深度神经网络对算力的需求不断增加,传统通用处理器在完成推理运算过程中出现了性能低、功耗高的缺点,因此通过专用硬件对深度神经网络进行加速逐步成为了深度神经网络的重要发展趋势。现场可编程门阵列(FPGA)具有重构性强、开发周期短以及性能优越等优点,适合用作深度神经网络的硬件加速平台。英伟达深度学习加速器(NVDLA)是英伟达开源的神经网络硬件加速器,其凭借自身出色的性能被学术界和工业界高度认可。本文主要研究NVDLA在FPGA平台上的优化映射问题,通过多种优化方案高效利用FPGA内部的硬件资源,同时提高其运行性能。基于搭建的NVDLA加速器平台,本文实现了对RESNET-50神经网络的硬件加速,完成了在ImageNet数据集上的图像分类任务。研究结果表明,优化后的NVDLA能显著提高硬件资源使用效率,处理性能最高可达30.8 fps,实现了较边缘中央处理器(CPU)加速器平台28倍的性能提升。     

基于深度学习的无线电信号对抗样本检测研究

针对无线电信号的攻击愈来愈频繁的情况,本文在数据流形理论基础上,使用深度神经网络(DNN)检测无线电信号对抗样本及其攻击方法。首先使用5种不同攻击方法对无线电信号进行攻击产生对抗样本,其次使用3种不同的神经网络检测对抗样本,最后用残差神经网络(ResNet)检测对抗样本的攻击方法。在信噪比(SNR)为30 d B和20 dB的无线电信号数据上的实验结果表明,本文所使用的残差神经网络检测精度接近100%,在信噪比为10 dB的无线电信号数据上的检测精度仍然在90%以上。结果表明本文所用的残差神经网络能有效检测无线电信号的对抗样本及其攻击方法。     

基于迁移学习与深度残差网络的滚动轴承快速故障诊断算法EI北大核心CSCD

针对现有基于深度学习的滚动轴承故障诊断算法训练参数量大,训练时间长且需要大量训练样本的缺点,提出了一种基于迁移学习(TL)与深度残差网络(ResNet)的快速故障诊断算法(TL-ResNet)。首先开发了一种将短时傅里叶变换(STFT)与伪彩色处理相结合的振动信号转三通道图像数据的方法;然后将在ImageNet数据集上训练的ResNet18模型作为预训练模型,通过迁移学习的方法,应用到滚动轴承故障诊断领域当中;最后对滚动轴承在不同工况下的故障诊断问题,提出了采用小样本迁移的方法进行诊断。在凯斯西储大学(CWRU)与帕德博恩大学(PU)数据集上进行了试验,TL-ResNet的诊断准确率分别为99.8%与95.2%,且在CWRU数据集上TL-ResNet的训练时间仅要1.5 s,这表明本算法优于其他的基于深度学习的故障诊断算法与经典算法,可用于实际工业环境中的快速故障诊断。     

基于变分模态分解与深度卷积神经网络的滚动轴承故障诊断

针对滚动轴承振动信号非平稳、非线性特点以及特征提取困难问题,提出一种基于变分模态分解(VMD)与深度卷积神经网络相结合的特征提取方法并应用于滚动轴承故障诊断。利用VMD将原始振动信号分解得到若干不同频率的限带本征模态分量,通过卷积网络中的多组卷积核自动学习各模态数据的不同特征,保证了特征提取的自适应性、全面性和多样性。在特征提取的基础上,使用全连接神经网络进行故障分类与诊断。将所提方法应用于滚动轴承故障诊断,结果表明,该方法在变工况情况下能够实现滚动轴承故障类别以及损伤程度的精确判定。     

基于小波散射卷积神经网络的结构损伤识别

损伤识别是结构状态评估领域的关键问题之一，对确保结构安全性有重要意义。深度学习算法在基于振动的结构损伤识别方面带来了许多突破，但从海量数据中挖掘结构损伤关键信息仍是亟待解决的技术难题。该研究提出了基于一维卷积神经网络(one-dimensional-convolutional neural network, 1D-CNN)深度学习的结构多类型损伤识别模型，采用小波散射变换对1D-CNN架构第一层卷积滤波器进行替换，通过散射系数实现输入层原始数据降维与特征提取，结合CNN卷积层、激活层和池化层实现监测数据特征增强处理。在此基础上，结合1D-CNN全连接层与Softmax函数实现特征数据分类，从而实现结构多类型损伤定位与定量高效识别。通过钢桁架结构和斜拉桥两种数值模型对上述框架进行了验证。结果表明：与普通卷积神经网络模型相比，基于小波散射卷积神经网络的结构损伤识别精度显著提升，损伤分类准确率达95.0%以上。随着传感数据环境噪声比例的增加，小波散射卷积神经网络损伤分类准确率虽略有下降，但仍保持较高精准度，说明该方法具有较强的鲁棒性抗噪能力。     

基于DTCNN-SVM的工业循环水系统供水泵故障诊断EI北大核心CSCD

工业循环水系统供水泵的工作状态是影响工业过程安全生产的重要因素,为及时准确地识别供水泵的工作状态,提出一种基于深度迁移卷积神经网络和支持向量机(deep transfer convolutional neural network-support vector machine,DTCNN-SVM)的故障诊断方法。将与工作状态强相关的振动信号进行信号-图像预处理,实现振动时序信号的二维灰度图化;在此基础上,采用融合迁移学习与残差神经网络的深度迁移卷积神经网络模型提取振动信号灰度图特征,并基于模糊不一致性度量对深度学习特征进行约简;采用支持向量机法建立供水泵故障诊断模型。试验结果表明,所提方法在少量样本数据和模型参数下能有效识别供水泵工作状态。     

基于特征融合进行活动识别的DCNN 5法

研究了输入是可穿戴传感器获得的多通道时间序列信号,输出是预定义的活动的活动识别模型,指出活动中的有效特征的提取目前多依赖于手工和浅层特征学习结构,不仅复杂而且会导致识别准确率下降;基于深度学习的卷积神经网络（ CNN）不是对时间序列信号进行手工特征提取,而是自动学习最优特征;目前使用卷积神经网络处理有限标签数据仍存在过拟合问题。因此提出了一种基于融合特征的系统性的特征学习方法用于活动识别,用ImageNet16对原始数据集进行预训练,将得到的数据与原始数据进行融合,并将融合数据和对应的标签送入有监督的深度卷积神经网络（ DCNN ）中,训练新的系统。在该系统中,特征学习和分类是相互加强的,它不仅能处理端到端的有限数据问题,也能使学习到的特征有更强的辨别力。与其他方法相比,该方法整体精度从87.0％提高到87.4％。     

基于MTF-CNN的轴承故障诊断研究EI北大核心CSCD

轴承故障诊断对保证机械设备的安全十分重要。近年来,数据驱动的故障诊断方法得到了研究者的关注。与传统的依赖于专家经验的故障特征提取方法不同,深度学习方法可以实现端到端自动故障特征提取与分类。针对一维信号作为卷积神经网络(convolutional neural network,CNN)输入时无法充分利用数据间的相关信息的问题,提出一种基于MTF-CNN的轴承故障诊断方法。利用马尔可夫变迁场(Markov transition field,MTF)对采集到的振动信号进行编码,根据数据之间的转移概率得到不同时间间隔内的数据相关性并生成相应特征图,之后将其输入卷积神经网络完成特征的提取并进行故障分类。采用凯斯西储大学轴承数据对模型进行验证,试验结果表明该模型达到99.8%以上的故障诊断准确率,与其他图像编码方式相比获得了较好的泛化性能。     

支持CNN与LSTM的二值权重神经网络芯片

深度神经网络在图像分类、语音识别、视频检测等领域都取得了巨大的成功,这些领域主要采用了卷积神经网络(CNN)、长短期记忆(LSTM)中的一种或者两种网络类型。由于CNN和LSTM网络结构的差异使得现有深度神经网络加速器无法同时高效支持这两种网络类型。权重二值化使得加速器对于CNN和LSTM的同时支持更加高效,同时使得计算复杂度和访存量大幅降低,使得神经网络加速器能够获得更高的能效,并且二值权重对中小规模神经网络模型的精度损失的影响非常有限。本文提出了一种高效支持CNN与LSTM的二值权重神经网络加速器设计结构,该结构在运行CNN和LSTM网络模型时,其核心运算单元利用率超过已有加速器,并且该加速器通过了片上系统(SoC)芯片验证,经过芯片实测,该加速器芯片能效在SoC系统级别达到了6.43 TOPS/W。     

基于深度神经网络的水声FBMC通信信号检测方法

针对传统水声滤波器组多载波（Filter Bank Multi-Carrier,FBMC）通信接收端需经过信道估计和均衡才可恢复出发送符号,系统复杂度高且信道估计精度不佳等问题。文章将深度神经网络融入到水声多载波通信当中,提出一种基于深度神经网络的水声FBMC信号检测方法。在训练阶段通过大量的数据迭代、调试超参数和优化算法来改善深度神经网络参数,使其具有预期的估计效果。利用训练完成的深度神经网络模型取代传统FBMC通信系统接收端的信道估计、均衡等模块,自适应地学习水声信道状态信息,同时避免了固有的虚部干扰影响。在测试阶段直接将频域序列作为网络的输入来预测发送的二进制序列,仿真结果表明所提出的基于深度神经网络的FBMC信号检测方法相比传统信道估计算法有更好的误码率性能。     

Investigation on the Chinese Text Sentiment Analysis Based on Convolutional Neural Networks in Deep Learning

Nowadays, the amount of wed data is increasing at a rapid speed, which presents a serious challenge to the web monitoring. Text sentiment analysis, an important research topic in the area of natural language processing, is a crucial task in the web monitoring area. The accuracy of traditional text sentiment analysis methods might be degraded in dealing with mass data. Deep learning is a hot research topic of the artificial intelligence in the recent years. By now, several research groups have studied the sentiment analysis of English texts using deep learning methods. In contrary, relatively few works have so far considered the Chinese text sentiment analysis toward this direction. In this paper, a method for analyzing the Chinese text sentiment is proposed based on the convolutional neural network (CNN) in deep learning in order to improve the analysis accuracy. The feature values of the CNN after the training process are nonuniformly distributed. In order to overcome this problem, a method for normalizing the feature values is proposed. Moreover, the dimensions of the text features are optimized through simulations. Finally, a method for updating the learning rate in the training process of the CNN is presented in order to achieve better performances. Experiment results on the typical datasets indicate that the accuracy of the proposed method can be improved compared with that of the traditional supervised machine learning methods, e.g., the support vector machine method.     

基于注意力机制的TDNN-LSTM模型及应用

在大数据规模下,基于深度学习的语音识别技术已经相当成熟,但在小样本资源下,由于特征信息的关联性有限,模型的上下文信息建模能力不足从而导致识别率不高。针对此问题,提出了一种嵌入注意力机制层（Attention Mechanism）的时延神经网络（Time Delay Neural Network,TDNN）结合长短时记忆递归（Long Short Term Memory,LSTM）神经网络的时序预测声学模型,即TLSTM-Attention,有效地融合了具有重要信息的粗细粒度特征以提高上下文信息建模能力。通过速度扰动技术扩增数据,结合说话人声道信息特征以及无词格最大互信息训练准则,选取不同输入特征、模型结构及节点个数进行对比实验。实验结果表明,该模型相比于基线模型,词错误率降低了3.37个百分点。     

Experimental exploration on loss surface of deep neural network

The loss function of the deep neural network is high dimensional, nonconvex and complex. So far, the geometric properties of the loss surface of the neural network have not been well understood. Different from most theoretical studies on the loss surface, this article makes the experimental exploration on the loss surface of the deep neural network, including trajectories of various adaptive optimization algorithms, the Hessian matrix of the loss function of the deep neural network, the curvature of the loss surface along the trajectories of the various adaptive optimization algorithms. It is found that the gradient direction of the adaptive optimization algorithms is almost perpendicular to the direction of the maximum curvature of the loss surface, while the gradient directions of the stochastic gradient descent (SGD) algorithm do not have such a rule. The Hessian matrix of the loss surface along the trajectory of the optimization algorithm is degraded, which is inconsistent with the hypothetical that nonsingular of the Hessian matrix in many theoretical studies of deep learning. Besides, this article proposes a new ensemble learning method of the neural network based on the scaling invariance of the ReLu neural network and mode connectivity.     

An Optimized Approach to Vehicle-Type Classification Using a Convolutional Neural Network

Vehicle type classification is considered a central part of an intelligent traffic system. In recent years, deep learning had a vital role in object detection in many computer vision tasks. To learn high-level deep features and semantics, deep learning offers powerful tools to address problems in traditional architectures of handcrafted feature-extraction techniques. Unlike other algorithms using handcrated visual features, convolutional neural network is able to automatically learn good features of vehicle type classification. This study develops an optimized automatic surveillance and auditing system to detect and classify vehicles of different categories. Transfer learning is used to quickly learn the features by recording a small number of training images from vehicle frontal view images. The proposed system employs extensive data-augmentation techniques for effective training while avoiding the problem of data shortage. In order to capture rich and discriminative information of vehicles, the convolutional neural network is fine-tuned for the classification of vehicle types using the augmented data. The network extracts the feature maps from the entire dataset and generates a label for each object (vehicle) in an image, which can help in vehicle-type detection and classification. Experimental results on a public dataset and our own dataset demonstrated that the proposed method is quite effective in detection and classification of different types of vehicles. The experimental results show that the proposed model achieves 96.04% accuracy on vehicle type classification.     

基于均值特征和改进深度神经网络的说话人识别算法

为提高神经网络在说话人识别应用中的识别性能,提出基于高斯增值矩阵特征和改进深度卷积神经网络的说话人识别算法。算法首先通过最大后验概率提取基于梅尔频率倒谱系数（Mel Frequency Cepstrum Coefficient,MFCC）特征的高斯均值矩阵,并对特征进行噪声适应性补偿,以增强信号的帧间关联和说话人特征信息,然后采用改进的深度卷积神经网络进一步对准帧间信息,以提高说话人识别特征对背景噪声的适应性。实验结果表明,相比于高斯混合模型-通用背景模型等识别框架及传统MFCC等特征,该算法可取得更高的识别准确率和最小的识别均方误差。     

Data Augmentation Using Contour Image for Convolutional Neural Network

With the development of artificial intelligence-related technologies such as deep learning, various organizations, including the government, are making various efforts to generate and manage big data for use in artificial intelligence. However, it is difficult to acquire big data due to various social problems and restrictions such as personal information leakage. There are many problems in introducing technology in fields that do not have enough training data necessary to apply deep learning technology. Therefore, this study proposes a mixed contour data augmentation technique, which is a data augmentation technique using contour images, to solve a problem caused by a lack of data. ResNet, a famous convolutional neural network (CNN) architecture, and CIFAR-10, a benchmark data set, are used for experimental performance evaluation to prove the superiority of the proposed method. And to prove that high performance improvement can be achieved even with a small training dataset, the ratio of the training dataset was divided into 70%, 50%, and 30% for comparative analysis. As a result of applying the mixed contour data augmentation technique, it was possible to achieve a classification accuracy improvement of up to 4.64% and high accuracy even with a small amount of data set. In addition, it is expected that the mixed contour data augmentation technique can be applied in various fields by proving the excellence of the proposed data augmentation technique using benchmark datasets.     

Autonomous Parking-Lots Detection with Multi-Sensor Data Fusion Using Machine Deep Learning Techniques

The rapid development and progress in deep machine-learning techniques have become a key factor in solving the future challenges of humanity. Vision-based target detection and object classification have been improved due to the development of deep learning algorithms. Data fusion in autonomous driving is a fact and a prerequisite task of data preprocessing from multi-sensors that provide a precise, well-engineered, and complete detection of objects, scene or events. The target of the current study is to develop an in-vehicle information system to prevent or at least mitigate traffic issues related to parking detection and traffic congestion detection. In this study we examined to solve these problems described by (1) extracting region-of-interest in the images (2) vehicle detection based on instance segmentation, and (3) building deep learning model based on the key features obtained from input parking images. We build a deep machine learning algorithm that enables collecting real video-camera feeds from vision sensors and predicting free parking spaces. Image augmentation techniques were performed using edge detection, cropping, refined by rotating, thresholding, resizing, or color augment to predict the region of bounding boxes. A deep convolutional neural network F-MTCNN model is proposed that simultaneously capable for compiling, training, validating and testing on parking video frames through video-camera. The results of proposed model employing on publicly available PK-Lot parking dataset and the optimized model achieved a relatively higher accuracy 97.6% than previous reported methodologies. Moreover, this article presents mathematical and simulation results using state-of-the-art deep learning technologies for smart parking space detection. The results are verified using Python, TensorFlow, OpenCV computer simulation frameworks.