基于RoBERTa和图增强Transformer的序列推荐方法

自推荐系统出现以来,有限的数据信息就一直制约着推荐算法的进一步发展。为降低数据稀疏性的影响,增强非评分数据的利用率,基于神经网络的文本推荐模型相继被提出,但主流的卷积或循环神经网络在文本语义理解和长距离关系捕捉方面存在明显劣势。为了更好地挖掘用户与商品之间的深层潜在特征,进一步提高推荐质量,提出一种基于Ro BERTa和图增强Transformer的序列推荐(RGT)模型。引入评论文本数据,首先利用预训练的Ro BERTa模型捕获评论文本中的字词语义特征,初步建模用户的个性化兴趣,然后根据用户与商品的历史交互信息,构建具有时序特性的商品关联图注意力机制网络模型,通过图增强Transformer的方法将图模型学习到的各个商品的特征表示以序列的形式输入Transformer编码层,最后将得到的输出向量与之前捕获的语义表征以及计算得到的商品关联图的全图表征输入全连接层,以捕获用户全局的兴趣偏好,实现用户对商品的预测评分。在3组真实亚马逊公开数据集上的实验结果表明,与Deep FM、Conv MF等经典文本推荐模型相比,RGT模型在均方根误差(RMSE)和平均绝对误差(MAE)2种指标上有显著提升,相较于最优对比模型最高分别提升4.7%和5.3%。

基于d-q变换及WOA-LSTM的异步电机定子匝间短路故障诊断方法

为了实现对异步电机定子绕组匝间短路故障的可靠在线诊断,提出一种基于d-q变换及鲸鱼优化算法(WOA)优化的长短期记忆网络(LSTM)的故障诊断方法。通过理论推导可知,d-q变换可有效提取定子电流中的特征频谱数据。采用鲸鱼优化算法对长短期记忆网络中的3个关键参数进行优化,建立WOA-LSTM故障分类模型。为了验证基于d-q变换和WOA-LSTM故障诊断方法的有效性,分别以小波变换、快速傅里叶变换及d-q变换提取电流频谱数据作为输入数据集,以一台YE2-100L1-4型异步电机为实验对象进行实验验证。研究结果表明:相比于小波变换及快速傅里叶变换,采用d-q变换能更准确的提取出定子电流中的故障特征,更精确地反映电机故障状态,有助于提高故障分类准确率;相比于传统的LSTM算法,经WOA优化后的LSTM算法分类准确率可达98.3%,能可靠地实现不同程度匝间短路故障的诊断。

基于IWOA-Transformer的磨煤机故障预警

提出了一种基于改进鲸鱼算法优化Transformer网络超参数(IWOA-Transformer)的故障预警方法。该方法利用非线性收敛系数和高斯变异对鲸鱼算法(WOA)进行改进,以提高WOA的收敛速度和避免其陷入局部最优;再采用改进鲸鱼算法(IWOA)优化Transformer的超参数,建立磨煤机故障预警模型;然后,通过预测值和实际值的相似度函数确定自适应阈值,结合专家系统判断故障类型并提出解决方案,实现磨煤机故障预警;最后,以某350 MW热电机组中速磨煤机为例进行故障预警试验。结果表明:所提IWOA-Transformer模型可显著提高预警速度和准确率,具有工程实用价值。

引入Transformer的道路小目标检测

针对道路场景中检测小目标时漏检率较高、检测精度低的问题,提出一种引入Transformer的道路小目标检测算法。在原YOLOv4算法基础上,对多尺度检测进行改进,把浅层特征信息充分利用起来;设计ICvT(improved convolutional vision transformer)模块捕获特征内部的相关性,获得上下文信息,提取更加全面丰富的特征;在网络特征融合部分嵌入改进后的空间金字塔池化模块,在保持较小计算量的同时增加特征图的感受野。实验结果表明,在KITTI数据集上,算法检测精度达到91.97%,与YOLOv4算法相比,mAP提高了2.53%,降低了小目标的漏检率。

基于TF-IDF和多头注意力Transformer模型的文本情感分析

文本情感分析旨在对带有情感色彩的主观性文本进行分析、处理、归纳和推理,是自然语言处理中一项重要任务。针对现有的计算方法不能充分处理复杂度和混淆度较高的文本数据集的问题,提出了一种基于TF-IDF(Term Frequency-Inverse Document Frequency)和多头注意力Transformer模型的文本情感分析模型。在文本预处理阶段,利用TF-IDF算法对影响文本情感倾向较大的词语进行初步筛选,舍去常见的停用词及其他文本所属邻域对文本情感倾向影响较小的专有名词。然后,利用多头注意力Transformer模型编码器进行特征提取,抓取文本内部重要的语义信息,提高模型对语义的分析和泛化能力。该模型在多领域、多类型评论语料库数据集上取得了98.17%的准确率。

Combined Economic and Emission Power Dispatch Control Using Substantial Augmented Transformative Algorithm

The purpose of the Combined Economic Emission Dispatch(CEED)of electric power is to offer the most exceptional schedule for production units,which must run with both low fuel costs and emission levels concurrently,thereby meeting the lack of system equality and inequality constraints.Economic and emissions dispatching has become a primary and significant concern in power system networks.Consequences of using non-renewable fuels as input to exhaust power systems with toxic gas emissions and depleted resources for future generations.The optimal power allocation to generators serves as a solution to this problem.Emission dispatch reduces emissions while ignoring economic considerations.A collective strategy known as Combined Economic and Emission Dispatch is utilized to resolve the above-mentioned problems and investigate the trade-off relationship between fuel cost and emissions.Consequently,this work manages the Substantial Augmented Transformative Algorithm(SATA)to take care of the Combined Economic Emission Dispatch Problem(CEEDP)of warm units while fulfilling imperatives,for example,confines on generator limit,diminish the fuel cost,lessen the emission and decrease the force misfortune.SATA is a stochastic streamlining process that relies upon the development and knowledge of swarms.The goal is to minimize the total fuel cost of fossil-based thermal power generation units that generate and cause environmental pollution.The algorithm searches for solutions in the search space from the smallest to the largest in the case of forwarding search.The simulation of the proposed system is developed using MATLAB Simulink software.Simulation results show the effectiveness and practicability of this method in terms of economic and emission dispatching issues.The performance of the proposed system is compared with existing Artificial Bee Colony-Particle Swarm Optimization(ABC-PSO),Simulated Annealing(SA),and Differential Evolution(DE)methods.The fuel cost and gas emission of the proposed system are 128904$/hr and 138094.4652$/hr.

Structural plane recognition from three-dimensional laser scanning points using an improved region-growing algorithm based on the robust randomized Hough transform

The staggered distribution of joints and fissures in space constitutes the weak part of any rock mass.The identification of rock mass structural planes and the extraction of characteristic parameters are the basis of rock-mass integrity evaluation,which is very important for analysis of slope stability.The laser scanning technique can be used to acquire the coordinate information pertaining to each point of the structural plane,but large amount of point cloud data,uneven density distribution,and noise point interference make the identification efficiency and accuracy of different types of structural planes limited by point cloud data analysis technology.A new point cloud identification and segmentation algorithm for rock mass structural surfaces is proposed.Based on the distribution states of the original point cloud in different neighborhoods in space,the point clouds are characterized by multi-dimensional eigenvalues and calculated by the robust randomized Hough transform(RRHT).The normal vector difference and the final eigenvalue are proposed for characteristic distinction,and the identification of rock mass structural surfaces is completed through regional growth,which strengthens the difference expression of point clouds.In addition,nearest Voxel downsampling is also introduced in the RRHT calculation,which further reduces the number of sources of neighborhood noises,thereby improving the accuracy and stability of the calculation.The advantages of the method have been verified by laboratory models.The results showed that the proposed method can better achieve the segmentation and statistics of structural planes with interfaces and sharp boundaries.The method works well in the identification of joints,fissures,and other structural planes on Mangshezhai slope in the Three Gorges Reservoir area,China.It can provide a stable and effective technique for the identification and segmentation of rock mass structural planes,which is beneficial in engineering practice.

基于改进YOLOv5s的CNN-Swin Transformer森林野生动物图像目标检测算法

【目的】为提高野生动物在复杂森林环境中的检测精度,促进森林野生动物保护技术发展,提出一种基于YOLOv5s网络模型、针对陷阱相机所摄取森林野生动物图像的改进检测算法。【方法】以包含湖南壶瓶山国家级自然保护区几种典型森林野生动物在内的数据集为研究对象,首先,对真实标注框图像进行裁剪、归一化和缩放处理,随机将2~4张裁剪图像拼贴组成新的数据集元素,以丰富和增强数据集图像信息;其次,使用一种基于通道注意力思想的加权通道拼接方法,在通道拼接时引入权重改变通道数量,通过反向传播训练方法不断更新权重以增加重要特征信息的通道层数;接着,引入Swin Transformer模块与CNN网络相结合,为卷积神经网络特征提取加入自注意力机制,融合2种网络特征提取层的优势,提高特征提取的感受野;最后,选择更优的α-DIoU损失函数替代GIoU损失函数,针对边界框重叠面积和中心点距离造成的损失,引入新的几何因素惩罚项。【结果】在相同试验条件和数据集下,相比原YOLOv5s网络模型,改进算法极大提高检测的平均准确率和平均回归率,均值平均精度由74.1%提升至88.4%,获得14.3%的精度提升,同时也超过YOLOv3、YOLOXs、RetinaNet、Faster R-CNN等其他流行目标检测算法。【结论】针对陷阱相机所摄取森林野生动物图像背景与目标对比度低、遮挡重叠严重,致使检测误检率、漏检率高等问题,在检测算法中提出一系列改进措施,为我国森林野生动物的保护和数据获取提供一种新的可行性方案和思路。

A Review of Image Steganography Based on Multiple Hashing Algorithm

Steganography is a technique for hiding secret messages while sending and receiving communications through a cover item.From ancient times to the present,the security of secret or vital information has always been a significant problem.The development of secure communication methods that keep recipient-only data transmissions secret has always been an area of interest.Therefore,several approaches,including steganography,have been developed by researchers over time to enable safe data transit.In this review,we have discussed image steganography based on Discrete Cosine Transform(DCT)algorithm,etc.We have also discussed image steganography based on multiple hashing algorithms like the Rivest–Shamir–Adleman(RSA)method,the Blowfish technique,and the hash-least significant bit(LSB)approach.In this review,a novel method of hiding information in images has been developed with minimal variance in image bits,making our method secure and effective.A cryptography mechanism was also used in this strategy.Before encoding the data and embedding it into a carry image,this review verifies that it has been encrypted.Usually,embedded text in photos conveys crucial signals about the content.This review employs hash table encryption on the message before hiding it within the picture to provide a more secure method of data transport.If the message is ever intercepted by a third party,there are several ways to stop this operation.A second level of security process implementation involves encrypting and decrypting steganography images using different hashing algorithms.

A New Image Watermarking Scheme Using Genetic Algorithm and Residual Numbers with Discrete Wavelet Transform

Transmission of data over the internet has become a critical issue as a result of the advancement in technology, since it is possible for pirates to steal the intellectual property of content owners. This paper presents a new digital watermarking scheme that combines some operators of the Genetic Algorithm (GA) and the Residue Number (RN) System (RNS) to perform encryption on an image, which is embedded into a cover image for the purposes of watermarking. Thus, an image watermarking scheme uses an encrypted image. The secret image is embedded in decomposed frames of the cover image achieved by applying a three-level Discrete Wavelet Transform (DWT). This is to ensure that the secret information is not exposed even when there is a successful attack on the cover information. Content creators can prove ownership of the multimedia content by unveiling the secret information in a court of law. The proposed scheme was tested with sample data using MATLAB2022 and the results of the simulation show a great deal of imperceptibility and robustness as compared to similar existing schemes.

Identification of Lubricating Oil Additives Using XGBoost and Ant Colony Optimization Algorithms

To address the problem of identifying multiple types of additives in lubricating oil,a method based on midinfrared spectral band selection using the eXtreme Gradient Boosting(XGBoost)algorithm combined with the ant colony optimization(ACO)algorithm is proposed.The XGBoost algorithm was used to train and test three additives,T534(alkyl diphenylamine),T308(isooctyl acid thiophospholipid octadecylamine),and T306(trimethylphenol phosphate),separately,in order to screen for the optimal combination of spectral bands for each additive.The ACO algorithm was used to optimize the parameters of the XGBoost algorithm to improve the identification accuracy.During this process,the support vector machine(SVM)and hybrid bat algorithms(HBA)were included as a comparison,generating four models:ACO-XGBoost,ACO-SVM,HBA-XGboost,and HBA-SVM.The results showed that all four models could identify the three additives efficiently,with the ACO-XGBoost model achieving 100%recognition of all three additives.In addition,the generalizability of the ACO-XGBoost model was further demonstrated by predicting a lubricating oil containing the three additives prepared in our laboratory and a collected sample of commercial oil currently in use。

基于LSTM-Transformer的城市轨道交通短时客流预测

准确预测城市轨道交通短时客流量的变化,有助于运营部门做出决策,并帮助轨道交通集团提高服务水平和实现智慧化运营。然而,客流数据的动态性和随机性使短时客流预测变得困难,因此,文章提出了一种组合预测模型,将Transformer模型中的位置编码(Positional Encoding)层与长短期记忆(Long Short-Term Memory,LSTM)神经网络相结合,构建了LSTM-Transformer预测模型。随后以青岛市的106个站点的进站客流数据为研究对象,并使用聚类算法对站点进行聚类分析。在10分钟的时间粒度下,利用前四周的客流数据作为训练数据,对未来一天的客流数据进行预测研究。同时,将差分自回归移动平均模型(Auto-Regressive Integrated Moving Average,ARIMA)、LSTM、GA-SLSTM和Transformer作为对照模型进行验证。通过多组实验证明了文章提出的LSTM-Transformer模型相较于对照模型组具有更好的预测精度和实用性。

基于Transformer的深度条件视频压缩

近年来,基于深度学习的视频压缩技术主要基于卷积神经网络(CNN)且采用运动补偿-残差编码的架构,由于常见的CNN只能利用局部的相关性,以及预测残差本身的稀疏特性,难以取得最优压缩性能。因此,提出一种基于Transformer架构的条件视频压缩算法,以实现更优的压缩效果。所提算法基于前后帧之间的运动信息,利用可形变卷积得到对应的预测帧特征;将预测帧特征作为条件信息,对原始输入帧特征进行条件编码,避免了直接编码稀疏的残差信号;利用特征间的非局部相关性,提出一个基于Transformer的深度条件视频压缩编码算法,用来实现运动信息编码和条件编码,进一步提升压缩编码的性能。实验结果表明:所提算法在HEVC、UVG数据集上均超越了当前主流的基于深度学习的视频压缩算法。

基于能量损失的Transformer神经网络信息流序列推荐算法

随着信息流和互联网的迅猛发展,网络越发成为人们获取信息的主要来源。有效提升用户浏览信息的效率,准确推送用户关注的个性化内容,成为当前的热门需求。利用Python爬取了平台一周时间内用户在信息流产品上的曝光历史,对数据进行处理和分析。引入Transformer深度神经网络模型和最相似用户估计模型并将其融合来预测用户浏览各个内容的点击率和浏览时长,模型解释性增强,且对不同顺序的推荐序列偏好更敏感。

融合Swin Transformer的YOLOv5口罩检测算法

针对口罩佩戴检测算法未平衡模型规模与检测精度的问题,提出了一种口罩佩戴检测改进算法。该算法以YOLOv5网络为基础框架:首先,应用轻量级Mixup数据增强方式和Mish激活函数以提高模型泛化能力;其次,引入Swin Transformer结构和ECA注意力机制来增强复杂场景下口罩目标的提取效率;第三,使用SIoU损失函数以提高检测精度;最后,设计了新的Neck网络卷积模块来实现模型轻量化。实验结果表明:相比于原始的YOLOv5算法,mAP提升2.9%,参数量减少54.2%,模型体积减少52.1%。该算法很好地平衡了模型规模与检测精度,在口罩检测实际场景中更具优势。

基于CEEMDAN-Transformer的灌浆流量混合预测模型

灌浆流量是最重要的水利工程灌浆参数之一,通过对灌浆流量的有效预测,可以实现对异常工况的提前响应,以保障施工质量与工程安全。然而由于灌浆过程面临的复杂地质情况,灌浆流量数据存在强非线性与波动性的特点,难以获得令人满意的计算精度。现有灌浆流量预测存在的不足如下:传统神经网络模型对时间序列特征提取和加工处理不足,导致预测精度有限;传统神经网络模型测试集进行一次计算仅能输出一个结果,进行多个时间步预测需要繁杂的多次计算;单测点预测结果预测时间短并且无法反映灌浆流量序列变化的整体趋势,不利于控制灌浆流量和保障施工质量。针对上述问题,本研究提出基于CEEMDAN-Transformer的灌浆流量混合预测模型。基于完全自适应噪声集合经验模态分解(Complete Ensemble Empirical Mode Decomposition with Adaptive Noise,CEEMDAN)方法将灌浆流量分解为本征模函数与残差信号,解决灌浆流量数据的非线性与强波动的问题;采用多头注意力Transformer实现多个本征模函数(Intrinsic Mode Function,IMF)序列到序列的预测,采用多头注意力机制来构建输入和输出的全局依赖关系,提升时间序列参数特征提取水平;最后,建立时序测点多输入多输出模型实现灌浆流量预测,提升多输出序列计算效率,反映整体趋势的多输出序列能够为灌浆流量控制提供参考。工程应用结果表明,本研究提出的基于CEEMDAN-Transformer的灌浆流量混合预测模型具有较好的计算精度和计算效率。

面向遥感目标检测的无锚框Transformer算法

遥感图像目标具有多方向排布、小且密集等特性,使基于深度学习的旋转目标检测算法存在检测精度不佳的问题.针对这一问题,本文提出了一种面向遥感目标检测的无锚框Transformer算法.首先,采用层次化Transformer采集不同分辨率的特征信息以扩大特征信息的采集范围.其次,构建一种新的前馈网络(Spacial-FeedForward Neural network,SFFN).SFFN将3×3深度可分离卷积的局部空间特性和多层感知机(MultiLayer Perceptron,MLP)的全局通道特性融合在一起,以解决前馈网络(Feed Forward Neural network,FFN)在局部空间建模上的不足.最后,基于SFFN架构搭建了无锚框检测器,将预测框回归问题分为水平框与旋转框,缓解了旋转框的损失不连续性问题.在DOTA数据集上的测试结果表明,此方法的平均精度达到了75.83%,同时在NWPU VHR-10数据集上5类小目标检测结果达到了92.47%,在遥感目标检测精度上更具竞争力.

Segmentation algorithm of complex ore images based on templates transformation and reconstruction

Lots of noises and heterogeneous objects with various sizes coexist in a complex image,such as an ore image;the classical image thresholding method cannot effectively distinguish between ores.To segment ore objects with various sizes simultaneously,two adaptive windows in the image were chosen for each pixel;the gray value of windows was calculated by Otsu＇s threshold method.To extract the object skeleton,the definition principle of distance transformation templates was proposed.The ores linked together in a binary image were separated by distance transformation and gray reconstruction.The seed region of each object was picked up from the local maximum gray region of the reconstruction image.Starting from these seed regions,the watershed method was used to segment ore object effectively.The proposed algorithm marks and segments most objects from complex images precisely.

Anti-aliasing nonstationary signals detecion algorithm based on interpolation in the frequency domain using the short time Fourier transform

To eliminate the aliasing that appeared during the measurement of multi-components nonstationary signals, a novel kind of anti-aliasing algorithm based on the short time Fourier transform （STFT） is brought forward. First the physical essence of aliasing that occurs is analyzed; second the interpolation algorithm model is setup based on the Hamming window; then the fast implementation of the algorithm using the Newton iteration method is given. Using the numerical simulation the feasibility of algorithm is validated. Finally, the electrical circuit experiment shows the practicality of the algorithm in the electrical engineering.

Spectral matching algorithm based on nonsubsampled contourlet transform and scale-invariant feature transform

A new spectral matching algorithm is proposed by us- ing nonsubsampled contourlet transform and scale-invariant fea- ture transform. The nonsubsampled contourlet transform is used to decompose an image into a low frequency image and several high frequency images, and the scale-invariant feature transform is employed to extract feature points from the low frequency im- age. A proximity matrix is constructed for the feature points of two related images. By singular value decomposition of the proximity matrix, a matching matrix （or matching result） reflecting the match- ing degree among feature points is obtained. Experimental results indicate that the proposed algorithm can reduce time complexity and possess a higher accuracy.