Convolutional neural networks(CNNs)are well suited to bearing fault classification due to their ability to learn discriminative spectro-temporal patterns.However,gathering sufficient cases of faulty conditions in real...Convolutional neural networks(CNNs)are well suited to bearing fault classification due to their ability to learn discriminative spectro-temporal patterns.However,gathering sufficient cases of faulty conditions in real-world engineering scenarios to train an intelligent diagnosis system is challenging.This paper proposes a fault diagnosis method combining several augmentation schemes to alleviate the problem of limited fault data.We begin by identifying relevant parameters that influence the construction of a spectrogram.We leverage the uncertainty principle in processing time-frequency domain signals,making it impossible to simultaneously achieve good time and frequency resolutions.A key determinant of this phenomenon is the window function's choice and length used in implementing the shorttime Fourier transform.The Gaussian,Kaiser,and rectangular windows are selected in the experimentation due to their diverse characteristics.The overlap parameter's size also influences the outcome and resolution of the spectrogram.A 50%overlap is used in the original data transformation,and±25%is used in implementing an effective augmentation policy to which two-stage regular CNN can be applied to achieve improved performance.The best model reaches an accuracy of 99.98%and a cross-domain accuracy of 92.54%.When combined with data augmentation,the proposed model yields cutting-edge results.展开更多
为加速AprilTag检测,提出了一种基于改进YOLOv5s预提取RoI(region of interest)的AprilTag检测方法。改进YOLOv5s网络,在输入灰度图像的单通道模式下,分别采用Ghost Bottleneck和ConvNeXt Block替换主干网络和颈部网络的C3和瓶颈模块,...为加速AprilTag检测,提出了一种基于改进YOLOv5s预提取RoI(region of interest)的AprilTag检测方法。改进YOLOv5s网络,在输入灰度图像的单通道模式下,分别采用Ghost Bottleneck和ConvNeXt Block替换主干网络和颈部网络的C3和瓶颈模块,提高模型的推理速度和泛化能力;通过亮度增强扩充数据集,提高模型鲁棒性。基于改进的YOLOv5网络进行AprilTag预识别,通过输出锚框划分RoI进行AprilTag检测,缩小图像处理范围,提高计算效率。实验结果表明,提出的AprilTag检测方法在1080P图像下FPS比传统AprilTag算法提高了77.42%以上。展开更多
针对数据存储中心硬盘故障数据稀少造成的故障预测效果不佳的问题,面向自我检测分析与报告技术(self-monitoring analysis and reporting technology,SMART)数据信息的时序特征,提出一种通过数据增强解决不平衡问题的硬盘故障预测算法...针对数据存储中心硬盘故障数据稀少造成的故障预测效果不佳的问题,面向自我检测分析与报告技术(self-monitoring analysis and reporting technology,SMART)数据信息的时序特征,提出一种通过数据增强解决不平衡问题的硬盘故障预测算法。该算法利用长短期记忆网络改进传统的生成对抗网络,生成包含故障恶化趋势信息的序列段数据,解决了数据集不平衡问题。同时,为进一步提高预测性能,预测模型融合了时序注意力机制和特征注意力机制,挖掘不同SMART特征和时间步对硬盘故障恶化过程的敏感程度。此外,在特征选择阶段结合了多种典型特征选择算法来选取关键特征。在真实硬盘数据集上进行了实验验证,结果表明,所提算法的准确率、召回率和F 1值均有较大提升。展开更多
针对大多数加密流量分类(encrypted traffic classification,ETC)模型由于标签数据稀缺而导致的性能下降问题,提出了一个基于对比学习的半监督加密流量分类(semisupervised encrypted traffic classification based on contrastive lear...针对大多数加密流量分类(encrypted traffic classification,ETC)模型由于标签数据稀缺而导致的性能下降问题,提出了一个基于对比学习的半监督加密流量分类(semisupervised encrypted traffic classification based on contrastive learning,SSETC-CL)模型。通过比较样本之间的相似性和差异性,SSETC-CL模型能够从大量无标注数据中学习到有用的表示,从而获得一个通用且优秀的特征编码网络,降低了下游任务对标签数据的依赖。本文在公有数据集ISCXVPN2016以及两个自采数据集上对SSETC-CL模型进行了评估。与其他基准模型相比,SSETC-CL模型在设定任务上的表现最佳,准确率最大提升了8.92%。实验结果表明,SSETC-CL模型不仅在预训练模型已知的流量上具有较高的精度,而且具备将预训练模型所获得的知识应用于未知流量的迁移能力。展开更多
基金supported by the National Natural Science Foundation of China(42027805)the National Aeronautical Fund(ASFC-20172080005)。
文摘Convolutional neural networks(CNNs)are well suited to bearing fault classification due to their ability to learn discriminative spectro-temporal patterns.However,gathering sufficient cases of faulty conditions in real-world engineering scenarios to train an intelligent diagnosis system is challenging.This paper proposes a fault diagnosis method combining several augmentation schemes to alleviate the problem of limited fault data.We begin by identifying relevant parameters that influence the construction of a spectrogram.We leverage the uncertainty principle in processing time-frequency domain signals,making it impossible to simultaneously achieve good time and frequency resolutions.A key determinant of this phenomenon is the window function's choice and length used in implementing the shorttime Fourier transform.The Gaussian,Kaiser,and rectangular windows are selected in the experimentation due to their diverse characteristics.The overlap parameter's size also influences the outcome and resolution of the spectrogram.A 50%overlap is used in the original data transformation,and±25%is used in implementing an effective augmentation policy to which two-stage regular CNN can be applied to achieve improved performance.The best model reaches an accuracy of 99.98%and a cross-domain accuracy of 92.54%.When combined with data augmentation,the proposed model yields cutting-edge results.
文摘为加速AprilTag检测,提出了一种基于改进YOLOv5s预提取RoI(region of interest)的AprilTag检测方法。改进YOLOv5s网络,在输入灰度图像的单通道模式下,分别采用Ghost Bottleneck和ConvNeXt Block替换主干网络和颈部网络的C3和瓶颈模块,提高模型的推理速度和泛化能力;通过亮度增强扩充数据集,提高模型鲁棒性。基于改进的YOLOv5网络进行AprilTag预识别,通过输出锚框划分RoI进行AprilTag检测,缩小图像处理范围,提高计算效率。实验结果表明,提出的AprilTag检测方法在1080P图像下FPS比传统AprilTag算法提高了77.42%以上。
文摘针对大多数加密流量分类(encrypted traffic classification,ETC)模型由于标签数据稀缺而导致的性能下降问题,提出了一个基于对比学习的半监督加密流量分类(semisupervised encrypted traffic classification based on contrastive learning,SSETC-CL)模型。通过比较样本之间的相似性和差异性,SSETC-CL模型能够从大量无标注数据中学习到有用的表示,从而获得一个通用且优秀的特征编码网络,降低了下游任务对标签数据的依赖。本文在公有数据集ISCXVPN2016以及两个自采数据集上对SSETC-CL模型进行了评估。与其他基准模型相比,SSETC-CL模型在设定任务上的表现最佳,准确率最大提升了8.92%。实验结果表明,SSETC-CL模型不仅在预训练模型已知的流量上具有较高的精度,而且具备将预训练模型所获得的知识应用于未知流量的迁移能力。
