基于深度学习的乳腺癌病理图像自动分类

乳腺癌病理图像的自动分类具有重要的临床应用价值。基于人工提取特征的分类算法,存在需要专业领域知识、耗时费力、提取高质量特征困难等问题。为此,采用一种改进的深度卷积神经网络模型,实现了乳腺癌病理图像的自动分类;同时,利用数据增强和迁移学习方法,有效避免了深度学习模型受样本量限制时易出现的过拟合问题。实验结果表明,该方法的识别率可达到91%,且具有较好的鲁棒性和泛化性。     

基于卷积神经网络的发动机气路故障诊断方法

深度学习是一种新的基于特征表示的机器学习方法。深度学习模型包含多个隐藏层,可以通过对输入数据进行自动学习来获取隐藏的功能层中的特征信息。与传统的诊断方法相比,深度学习具备从原始信息中提取更丰富的特征的能力,因此已经成为基于机器学习的故障诊断研究的新方向,为发动机气路等复杂系统故障诊断带来了新思路。结合发动机气路试验数据的特点与深度学习的优势,提出基于卷积神经网络的故障诊断方法,包括预处理、模型训练及优化等过程,并实现了复杂系统故障诊断预测算法平台。经某发动机气路试验仿真数据实例验证,提出的方法具有较好的可行性和效果,能够充分利用深度学习的优点,更准确地识别发动机气路的健康状况。     

基于堆叠降噪稀疏自动编码器的软件缺陷预测

特征提取是软件缺陷预测中的关键步骤,特征提取的质量决定了缺陷预测模型的性能,但传统的特征提取方法难以提取出软件缺陷数据的深层本质特征。深度学习理论中的自动编码器能够从原始数据中自动学习特征,并获得其特征表示,同时为了增强自动编码器的鲁棒性,本文提出一种基于堆叠降噪稀疏自动编码器的特征提取方法,通过设置不同的隐藏层数、稀疏性约束和加噪方式,可以直接高效地从软件缺陷数据中提取出分类预测所需的各层次特征表示。利用Eclipse缺陷数据集的实验结果表明,该方法较传统特征提取方法具有更好的性能。     

基于小样本学习的SAR图像识别

深度学习已成为图像识别领域的一个研究热点。与传统图像识别方法不同,深度学习从大量数据中自动学习特征,并且具有强大的自学习能力和高效的特征表达能力。但在小样本条件下,传统的深度学习方法如卷积神经网络难以学习到有效的特征,造成图像识别的准确率较低。因此,提出一种新的小样本条件下的图像识别算法用于解决SAR图像的分类识别。该算法以卷积神经网络为基础,结合自编码器,形成深度卷积自编码网络结构。首先对图像进行预处理,使用2D Gabor滤波增强图像,在此基础上对模型进行训练,最后构建图像分类模型。该算法设计的网络结构能自动学习并提取小样本图像中的有效特征,进而提高识别准确率。在MSTAR数据集的10类目标分类中,选择训练集数据中10%的样本作为新的训练数据,其余数据为验证数据,并且,测试数据在卷积神经网络中的识别准确率为76.38%,而在提出的卷积自编码结构中的识别准确率达到了88.09%。实验结果表明,提出的算法在小样本图像识别中比卷积神经网络模型更加有效。     

基于并联CNN-SE-Bi-LSTM的轴承剩余使用寿命预测

滚动轴承作为一种机械标准件,广泛应用于各类旋转机械设备,其健康状况对机器设备的正常运行至关重要,掌握其剩余使用寿命(RUL)可以更好地保证生产活动安全有效的进行.针对目前基于深度学习的机器RUL预测方法普遍存在:a)预测性能很大程度依赖手工特征设计;b)模型不能够充分提取数据中的有用特征;c)学习过程中没有明确考虑多传感器数据等缺点,提出了一种新的深度预测网络——并联多个带有压缩激励机制的卷积神经网络和双向长短期记忆网络集成网络(CNN-SE-Bi-LSTM),用于设备的RUL预测.在该预测网络中,不同传感器采集的监测数据直接作为预测网络的输入.然后,在改进的压缩激励卷积网络(CNN-SE-Net)提取空间特征的基础上进一步通过双向长短期记忆网络(Bi-LSTM)提取时序特征,建立起多个独立的可以自动从输入数据中学习高级表示的RU L预测模型分支.最后,将各独立分支学习到的特征通过全连接层并联获得最终的RU L预测模型.通过滚动轴承加速退化实验的数据,验证了所提网络的有效性并与现有的一些改进算法进行了对比实验.结果表明,面对原始多传感器数据,该算法能够自适应地提供准确的RU L预测结果,且预测表现优于现有一些预测方法.     

基于原型学习与深度特征融合的脑功能连接分类方法研究EI北大核心CSCD

近年来,基于深度学习的脑功能连接分类方法已成为一个研究热点.为了进一步提高脑功能连接的分类准确率,获得与疾病相关的鉴别性特征,本文提出了一种基于原型学习与深度特征融合的脑功能连接分类方法.该方法首先使用栈式自编码器从脑功能连接中提取从低层次到高层次的深度特征;然后利用原型学习在自编码器的各隐层中提取表示样本类别信息的距离特征;最后采用深度特征融合策略将这些距离特征融合,并将该融合特征用于脑功能连接的类别标签预测.在ABIDE数据集上的实验结果表明,与其他同类方法相比,该方法不仅具有较高的分类准确率,而且能够更加准确地定位与疾病相关的脑区.     

基于深度多特征融合的CNNs图像分类算法

近年来,卷积神经网络在图像处理方面的良好性能得到了广泛关注.为了更好地提取图像内容信息,提高图像分类精度,提出了一种基于深度多特征融合的CNNs图像分类算法.算法有效深度融合了图像的多种特征,即使用k-means++聚类算法提取的主颜色特征和利用去噪卷积神经网络提取的空间位置特征.实验结果表明,提出的基于深度多特征融合...     

基于特征迁移学习的综合能源系统小样本日前电力负荷预测

基于历史数据和深度学习的负荷预测已广泛应用于以电能为中心的综合能源系统中以提高预测精度,然而,当区域中出现新用户时,其历史负荷数据往往极少,此时,深度学习难以适用.针对此,本文提出基于负荷特征提取和迁移学习的预测机制.首先,依据源域用户历史负荷数据,融合聚类算法和门控循环单元网络构建源域数据的特征提取和分类模型;然后,利用该模型提取当前待预测目标域小样本的特征及其类别信息,进而给出基于特征相似度和时间遗忘因子的特征融合策略;最后,依据融合特征,给出基于迁移学习和特征输入的负荷预测.将所提算法应用于卡迪夫某区域的高中和住宅用电预测中,实验结果表明了该算法在综合能源系统小样本电力负荷预测中的有效性.     

基于增强深度特征和TSK模糊分类器的癫痫脑电信号识别

识别癫痫脑电信号的关键在于获取有效的特征和构建可解释的分类器.为此,提出一种基于增强深度特征的TSK模糊分类器(ED-TSK-FC).首先,ED-TSK-FC使用一维卷积神经网络(1D-CNN)自动获取癫痫脑电信号的深度特征与潜在类别信息,并将深度特征和潜在类别信息合并为增强深度特征;其次,将增强深度特征作为ED-TSK-FC模糊规则前件与后件部分的训练变量,保证原始输入的深度特征及其潜在意义都出现在模糊规则中,进而对增强深度特征作出良好的解释;然后,采用岭回归极限学习算法对模糊规则的后件参数进行快速求解,在不显著降低分类准确度的情况下,ED-TSK-FC的廉价训练方法可以缩短模型的训练时间;最后,在Bonn癫痫数据集上,分别从分类性能、学习效率和可解释性3个方面,验证ED-TSK-FC的优越性.     

基于深度学习的脸部年龄预测算法

为提高脸部年龄预测的准确性,在深度学习的基础上提出一种可有效预测脸部年龄的算法。通过对人脸图像进行预处理,获取左眼、右眼、鼻子和嘴巴四个部分的局部图像,利用迁移TensorFlow深度学习库中的Inception V4模型,提取脸部图像四个部分的多尺度局部特征,并将提取的局部特征使用串联方式相连接以得到融合特征,再将不同年龄的融合特征输入双向长短期记忆中,以学习不同年龄融合特征间的相关性,进而完成年龄预测。在公开数据集FG-NET和MORPH上的实验结果表明,该算法通过利用脸部多尺度融合特征和不同年龄融合特征间的相关性,能够显著提高年龄预测的准确性和鲁棒性。     

Distress classification of class-imbalanced inspection data via correlation-maximizing weighted extreme learning machine

This paper presents distress classification of class-imbalanced inspection data via correlation-maximizing weighted extreme learning machine (CMWELM). For distress classification, it is necessary to extract semantic features that can effectively distinguish multiple kinds of distress from a small amount of class-imbalanced data. In recent machine learning techniques such as general deep learning methods, since effective feature transformation from visual features to semantic features can be realized by using multiple hidden layers, a large amount of training data are required. However, since the amount of training data of civil structures becomes small, it becomes difficult to perform successful transformation by using these multiple hidden layers. On the other hand, CMWELM consists of two hidden layers. The first hidden layer performs feature transformation, which can directly extract the semantic features from visual features, and the second hidden layer performs classification with solving the class-imbalanced problem. Specifically, in the first hidden layer, the feature transformation is realized by using projections obtained by maximizing the canonical correlation between visual and text features as weight parameters of the hidden layer without designing multiple hidden layers. Furthermore, the second hidden layer enables successful training of our classifier by using weighting factors concerning the class-imbalanced problem. Consequently, CMWELM realizes accurate distress classification from a small amount of class-imbalanced data.     

Times-series data augmentation and deep learning for construction equipment activity recognition

Automated, real-time, and reliable equipment activity recognition on construction sites can help to minimize idle time, improve operational efficiency, and reduce emissions. Previous efforts in activity recognition of construction equipment have explored different classification algorithms anm accelerometers and gyroscopes. These studies utilized pattern recognition approaches such as statistical models (e.g., hidden-Markov models); shallow neural networks (e.g., Artificial Neural Networks); and distance algorithms (e.g., K-nearest neighbor) to classify the time-series data collected from sensors mounted on the equipment. Such methods necessitate the segmentation of continuous operational data with fixed or dynamic windows to extract statistical features. This heuristic and manual feature extraction process is limited by human knowledge and can only extract human-specified shallow features. However, recent developments in deep neural networks, specifically recurrent neural network (RNN), presents new opportunities to classify sequential time-series data with recurrent lateral connections. RNN can automatically learn high-level representative features through the network instead of being manually designed, making it more suitable for complex activity recognition. However, the application of RNN requires a large training dataset which poses a practical challenge to obtain from real construction sites. Thus, this study presents a data-augmentation framework for generating synthetic time-series training data for an RNN-based deep learning network to accurately and reliably recognize equipment activities. The proposed methodology is validated by generating synthetic data from sample datasets, that were collected from two earthmoving operations in the real world. The synthetic data along with the collected data were used to train a long short-term memory (LSTM)-based RNN. The trained model was evaluated by comparing its performance with traditionally used classification algorithms for construction equipment activity recognition. The deep learning framework presented in this study outperformed the traditionally used machine learning classification algorithms for activity recognition regarding model accuracy and generalization.     

Deep learning for fault-relevant feature extraction and fault classification with stacked supervised auto-encoder

Stacked auto-encoder (SAE)-based deep learning has been introduced for fault classification in recent years, which has the potential to extract deep abstract features from the raw input data. However, SAE cannot ensure the relevance of deep features with the fault types due to its unsupervised self-reconstruction in the pretraining stage. To overcome this problem, a stacked supervised auto-encoder is proposed to pretrain the deep network and obtain deep fault-relevant features from raw input data. In each supervised auto-encoder, informative features are learned from the input data with the goal that they can largely distinguish different fault types. By stacking multiple supervised auto-encoders hierarchically, high-level fault-relevant features are gradually learned from raw input data, which can improve the classification accuracy of the classifiers. The proposed SSAE is tested on the Tennessee–Eastman (TE) benchmark process and a real industrial hydrocracking process. The results show the effectiveness and flexibility of SSAE.     

多粒度融合的模糊规则系统图像特征学习

当前最流行的图像特征学习方法是深度神经网络,该类方法无需人工参与即可自动地通过特征学习提取高效的特征,用于分类识别等任务。然而,深度神经网络图像特征抽取方法目前也面临着诸多挑战,其有效性严重依赖大规模的数据,且通常被视为黑盒模型,解释性较差。针对上述挑战,以基于模糊规则推理的TSK模糊系统(TSK-FS)为基础,提出了一种适用于不同规模数据集且易于理解的特征学习方法——多粒度融合的模糊规则系统图像特征学习算法。该方法通过基于规则的TSK-FS抽取图像特征,因而特征学习过程是可以利用规则进行解释的。其次,多粒度扫描也使得其特征学习能力进一步提升。在不同规模的图像数据集上进行了充分的实验,实验结果表明该方法在图像数据集上具有较好的有效性。     

基于变分自编码器的谣言立场分类算法

针对当前谣言检测任务中社交媒体推特平台的推文数据分布复杂且不均衡的特点,提出基于变分自编码器（VAE）的谣言立场分类算法VAE-LSTM。对数据进行预处理后,利用word2vec模型提取推文词向量并输入VAE中进行训练,得到符合简单概率分布的深度特征序列再从中采样获取有效特征,以避免数据量较大的推文类别影响特征向量。在此基础上,使用长短时记忆（LSTM）网络处理向量序列数据进而实现分类。理论分析和实验结果表明,VAE-LSTM算法无须手动提取或添加特征,训练过程简单高效,同时能缓解类间不平衡问题,其应用于实际场景准确率和F1得分分别为0.800和0.494,与时序注意力机制算法、Turing算法、霍克斯过程算法等相比分类性能更好,且较SVM等早期机器学习方法节省了大量数据预处理时间。     

On combining multiscale deep learning features for the classification of hyperspectral remote sensing imagery

In recent years, satellite imagery has greatly improved in both spatial and spectral resolution. One of the major unsolved problems in highly developed remote sensing imagery is the manual selection and combination of appropriate features according to spectral and spatial properties. Deep learning framework can learn global and robust features from the training data set automatically, and it has achieved state-of-the-art classification accuracies over different image classification tasks. In this study, a technique is proposed which attempts to classify hyperspectral imagery by incorporating deep learning features. Firstly, deep learning features are extracted by multiscale convolutional auto-encoder. Then, based on the learned deep learning features, a logistic regression classifier is trained for classification. Finally, parameters of deep learning framework are analysed and the potential development is introduced. Experiments are conducted on the well-known Pavia data set which is acquired by the reflective optics system imaging spectrometer sensor. It is found that the deep learning-based method provides a more accurate classification result than the traditional ones.