Lightweight Underwater Target Detection Using YOLOv8 with Multi-Scale Cross-Channel Attention

Underwater target detection is extensively applied in domains such as underwater search and rescue,environmental monitoring,and marine resource surveys.It is crucial in enabling autonomous underwater robot operations and promoting ocean exploration.Nevertheless,low imaging quality,harsh underwater environments,and obscured objects considerably increase the difficulty of detecting underwater targets,making it difficult for current detection methods to achieve optimal performance.In order to enhance underwater object perception and improve target detection precision,we propose a lightweight underwater target detection method using You Only Look Once(YOLO)v8 with multi-scale cross-channel attention(MSCCA),named YOLOv8-UOD.In the proposed multiscale cross-channel attention module,multi-scale attention(MSA)augments the variety of attentional perception by extracting information from innately diverse sensory fields.The cross-channel strategy utilizes RepVGGbased channel shuffling(RCS)and one-shot aggregation(OSA)to rearrange feature map channels according to specific rules.It aggregates all features only once in the final feature mapping,resulting in the extraction of more comprehensive and valuable feature information.The experimental results show that the proposed YOLOv8-UOD achieves a mAP50 of 95.67%and FLOPs of 23.8 G on the Underwater Robot Picking Contest 2017(URPC2017)dataset,outperforming other methods in terms of detection precision and computational cost-efficiency.

Engine Misfire Fault Detection Based on the Channel Attention Convolutional Model

To accurately diagnosemisfire faults in automotive engines,we propose a Channel Attention Convolutional Model,specifically the Squeeze-and-Excitation Networks(SENET),for classifying engine vibration signals and precisely pinpointing misfire faults.In the experiment,we established a total of 11 distinct states,encompassing the engine’s normal state,single-cylinder misfire faults,and dual-cylinder misfire faults for different cylinders.Data collection was facilitated by a highly sensitive acceleration signal collector with a high sampling rate of 20,840Hz.The collected data were methodically divided into training and testing sets based on different experimental groups to ensure generalization and prevent overlap between the two sets.The results revealed that,with a vibration acceleration sequence of 1000 time steps(approximately 50 ms)as input,the SENET model achieved a misfire fault detection accuracy of 99.8%.For comparison,we also trained and tested several commonly used models,including Long Short-Term Memory(LSTM),Transformer,and Multi-Scale Residual Networks(MSRESNET),yielding accuracy rates of 84%,79%,and 95%,respectively.This underscores the superior accuracy of the SENET model in detecting engine misfire faults compared to other models.Furthermore,the F1 scores for each type of recognition in the SENET model surpassed 0.98,outperforming the baseline models.Our analysis indicated that the misclassified samples in the LSTM and Transformer models’predictions were primarily due to intra-class misidentifications between single-cylinder and dual-cylinder misfire scenarios.To delve deeper,we conducted a visual analysis of the features extracted by the LSTM and SENET models using T-distributed Stochastic Neighbor Embedding(T-SNE)technology.The findings revealed that,in the LSTMmodel,data points of the same type tended to cluster together with significant overlap.Conversely,in the SENET model,data points of various types were more widely and evenly dispersed,demonstrating its effectiveness in distinguishing between different fault types.

Attention Eraser and Quantitative Measures for Automated Bone Age Assessment

Bone age assessment(BAA)aims to determine whether a child’s growth and development are normal concerning their chronological age.To predict bone age more accurately based on radiographs,and for the left-hand X-ray images of different races model can have better adaptability,we propose a neural network in parallel with the quantitative features from the left-hand bone measurements for BAA.In this study,a lightweight feature extractor(LFE)is designed to obtain the featuremaps fromradiographs,and amodule called attention erasermodule(AEM)is proposed to capture the fine-grained features.Meanwhile,the dimensional information of the metacarpal parts in the radiographs is measured to enhance the model’s generalization capability across images fromdifferent races.Ourmodel is trained and validated on the RSNA,RHPE,and digital hand atlas datasets,which include images from various racial groups.The model achieves a mean absolute error(MAE)of 4.42 months on the RSNA dataset and 15.98 months on the RHPE dataset.Compared to ResNet50,InceptionV3,and several state-of-the-art methods,our proposed method shows statistically significant improvements(p<0.05),with a reduction in MAE by 0.2±0.02 years across different racial datasets.Furthermore,t-tests on the features also confirm the statistical significance of our approach(p<0.05).

Research on Stock Price Prediction Method Based on the GAN-LSTM-Attention Model

Stock price prediction is a typical complex time series prediction problem characterized by dynamics,nonlinearity,and complexity.This paper introduces a generative adversarial network model that incorporates an attention mechanism(GAN-LSTM-Attention)to improve the accuracy of stock price prediction.Firstly,the generator of this model combines the Long and Short-Term Memory Network(LSTM),the Attention Mechanism and,the Fully-Connected Layer,focusing on generating the predicted stock price.The discriminator combines the Convolutional Neural Network(CNN)and the Fully-Connected Layer to discriminate between real stock prices and generated stock prices.Secondly,to evaluate the practical application ability and generalization ability of the GAN-LSTM-Attention model,four representative stocks in the United States of America(USA)stock market,namely,Standard&Poor’s 500 Index stock,Apple Incorporatedstock,AdvancedMicroDevices Incorporatedstock,and Google Incorporated stock were selected for prediction experiments,and the prediction performance was comprehensively evaluated by using the three evaluation metrics,namely,mean absolute error(MAE),root mean square error(RMSE),and coefficient of determination(R2).Finally,the specific effects of the attention mechanism,convolutional layer,and fully-connected layer on the prediction performance of the model are systematically analyzed through ablation study.The results of experiment show that the GAN-LSTM-Attention model exhibits excellent performance and robustness in stock price prediction.

Application of the improved dung beetle optimizer,muti-head attention and hybrid deep learning algorithms to groundwater depth prediction in the Ningxia area,China

Due to the lack of accurate data and complex parameterization,the prediction of groundwater depth is a chal-lenge for numerical models.Machine learning can effectively solve this issue and has been proven useful in the prediction of groundwater depth in many areas.In this study,two new models are applied to the prediction of groundwater depth in the Ningxia area,China.The two models combine the improved dung beetle optimizer(DBO)algorithm with two deep learning models:The Multi-head Attention-Convolution Neural Network-Long Short Term Memory networks(MH-CNN-LSTM)and the Multi-head Attention-Convolution Neural Network-Gated Recurrent Unit(MH-CNN-GRU).The models with DBO show better prediction performance,with larger R(correlation coefficient),RPD(residual prediction deviation),and lower RMSE(root-mean-square error).Com-pared with the models with the original DBO,the R and RPD of models with the improved DBO increase by over 1.5%,and the RMSE decreases by over 1.8%,indicating better prediction results.In addition,compared with the multiple linear regression model,a traditional statistical model,deep learning models have better prediction performance.

The Impairment Attention Capture by Topological Change in Children With Autism Spectrum Disorder

Objective Autism spectrum disorder(ASD)is a neurodevelopmental condition characterized by difficulties with communication and social interaction,restricted and repetitive behaviors.Previous studies have indicated that individuals with ASD exhibit early and lifelong attention deficits,which are closely related to the core symptoms of ASD.Basic visual attention processes may provide a critical foundation for their social communication and interaction abilities.Therefore,this study explores the behavior of children with ASD in capturing attention to changes in topological properties.Methods Our study recruited twenty-seven ASD children diagnosed by professional clinicians according to DSM-5 and twenty-eight typically developing(TD)age-matched controls.In an attention capture task,we recorded the saccadic behaviors of children with ASD and TD in response to topological change(TC)and non-topological change(nTC)stimuli.Saccadic reaction time(SRT),visual search time(VS),and first fixation dwell time(FFDT)were used as indicators of attentional bias.Pearson correlation tests between the clinical assessment scales and attentional bias were conducted.Results This study found that TD children had significantly faster SRT(P<0.05)and VS(P<0.05)for the TC stimuli compared to the nTC stimuli,while the children with ASD did not exhibit significant differences in either measure(P>0.05).Additionally,ASD children demonstrated significantly less attention towards the TC targets(measured by FFDT),in comparison to TD children(P<0.05).Furthermore,ASD children exhibited a significant negative linear correlation between their attentional bias(measured by VS)and their scores on the compulsive subscale(P<0.05).Conclusion The results suggest that children with ASD have difficulty shifting their attention to objects with topological changes during change detection.This atypical attention may affect the child’s cognitive and behavioral development,thereby impacting their social communication and interaction.In sum,our findings indicate that difficulties in attentional capture by TC may be a key feature of ASD.

基于Attention-LSTM的短期电力负荷预测

电力负荷预测的准确性受到多种因素的干扰,如气候变化、经济发展以及区域差异等,这些因素使得电力负荷呈现出显著的不稳定性和复杂的非线性特征,从而增加了提高预测精度的难度。为了应对这一挑战,本文创新性地引入了一种结合自注意力机制与长短期记忆网络(LSTM)的预测方法。通过在美国某一地区的实际用电负荷数据验证模型,实验结果表明,该方法的决定系数(R2)为0.96,平均绝对误差(MAE)为0.023,均方根误差(RMSE)为0.029,提升了预测的准确性。这不仅证明了所提模型在提高电力负荷预测精度方面的有效性,也为其在船舶电力负荷预测的应用奠定了一定的基础。

引入上下文信息和Attention Gate的GUS-YOLO遥感目标检测算法

目前基于通用YOLO系列的遥感目标检测算法存在并未充分利用图像的全局上下文信息,在特征融合金字塔部分并未充分考虑缩小融合特征之间的语义鸿沟、抑制冗余信息干扰的缺点。在结合YOLO算法优点的基础上提出GUS-YOLO算法,其拥有一个能够充分利用全局上下文信息的骨干网络Global Backbone。除此之外,该算法在融合特征金字塔自顶向下的结构中引入Attention Gate模块,可以突出必要的特征信息,抑制冗余信息。另外,为Attention Gate模块设计了最佳的网络结构,提出了网络的特征融合结构U-Net。最后,为克服ReLU函数可能导致模型梯度不再更新的问题,该算法将Attention Gate模块的激活函数升级为可学习的SMU激活函数,提高模型鲁棒性。在NWPU VHR-10遥感数据集上,该算法相较于YOLOV7算法取得宽松指标mAP^(0.50)1.64个百分点和严格指标mAP^(0.75)9.39个百分点的性能提升。相较于目前主流的七种检测算法,该算法取得较好的检测性能。

基于IGWO-Attention-GRU的短期电力负荷预测模型

为了提高短期电力负荷的预测精度,针对电力负荷序列波动性强、复杂性高的特点,综合考虑气象因素及日期类型的影响,文中提出一种基于改进灰狼优化算法(IGWO)优化Attention-GRU网络的短期电力负荷预测模型。首先,构建Attention-GRU网络;其次,对灰狼优化算法(GWO)进行改进,并利用IGWO寻找Attention-GRU网络的超参数;最后,使用IGWO-Attention-GRU模型在电力负荷数据集上进行实验,并与多种预测模型进行比较。实验结果表明,IGWO-Attention-GRU模型的MAPE、RMSE和MAE值均为各种预测模型中最低,验证了IGWO-Attention-GRU模型的优越性。

基于SABO-GRU-Attention的锂电池SOC估计

提出一种基于SABO-GRU-Attention(subtraction average based optimizer-gate recurrent unitattention)的锂电池SOC(state of charge)估计方法。采用基于平均减法优化算法自适应更新GRU神经网络的超参数,融合SE(squeeze and excitation)注意力机制自适应分配各通道权重,提高学习效率。对马里兰大学电池数据集进行预处理,输入电压、电流参数,进行锂电池充放电仿真实验,并搭建锂电池荷电状态实验平台进行储能锂电池充放电实验。结果表明,提出的SOC神经网络估计模型明显优于LSTM、GRU以及PSO-GRU等模型,具有较高的估计精度与应用价值。

血浆LPA、TGF⁃β1和D⁃D检测对全膝关节置换术后DVT诊断的价值

目的分析血浆血磷脂酸(LPA)、转化生长因子β1(TGF⁃β1)和D⁃二聚体(D⁃D)检测对全膝关节置换术后下肢深静脉血栓(DVT)诊断的价值。方法回顾性分析2021年5月至2023年10月于山西省荣军医院行全膝关节置换术的89例患者的临床资料,按照术后彩色多普勒超声结果分为非DVT组(n=69)和DVT组(n=20)。比较两组血浆LPA、TGF⁃β1和D⁃D水平,比较两组组临床资料(年龄、性别、运动、吸烟史、BMI、家族史、采取血栓预防措施),采用多元Logistic回归分析影响影响全膝关节置换术后DV发生的危险因素,采用ROC曲线分析血浆LPA、TGF⁃β1和D⁃D单一、联合对全膝关节置换术后DVT的诊断价值。结果DVT组血浆LPA、TGF⁃β1和D⁃D均显著高于非DVT组,差异有统计学意义(P<0.05);两组年龄、性别、运动、吸烟史、BMI和家族史比较差异无统计学意义(P>0.05)。两组的采取血栓预防措施比较差异有统计学意义(P<0.05);多元Logistic回归分析结果显示:未采取血栓预防措施、血浆LPA(<10μmol/L,>140μmol/L)、TGF⁃β1(<10 pg/mL,>100 pg/mL)、D⁃D(>0.5 mg/L)均为影响全膝关节置换术后DV发生的危险因素(P<0.05);血浆LPA+TGF⁃β1+D⁃D联合诊断全膝关节置换术后DVT的AUC为0.896,高于血浆LPA、TGF⁃β1和D⁃D单独检测(P<0.05)。结论血浆LPA、TGF⁃β1、D⁃D水平在全膝关节置换术后DVT发生患者中呈高表达,且三指标联合检测可提高全膝关节置换术后DVT的诊断效果。

基于XGBoost-WOA-BiLSTM-Attention的公共建筑暖通空调能耗预测研究

为在双碳目标下实现节能减排,降低能源成本,提出一种基于BiLSTM的公共建筑暖通空调能耗预测模型。在BiLSTM模型基础上,使用XGBoost算法对输入特征进行选择,剔除冗余特征,得到最佳模型输入特征;然后利用WOA优化算法对添加了Attention机制的BiLSTM模型中的6个超参数进行优化,将得到的最优参数代入BiLSTM-Attention神经网络中进行预测,并与BiLSTM模型、BiLSTM-Attention模型和WOA-BiLSTM-Attention模型进行对比。结果表明,所提出的XGBoost-WOA-BiLSTM-Attention模型的RMSE、MAE、R2分别为0.0106、0.006、0.9991,优于其他模型,且相对于持续模型在均方根误差RMSE上提升了98%,为降低公共建筑暖通空调能耗研究提供了参考。

基于CNN-BiGRU-Attention的短期电力负荷预测

针对目前电力负荷数据随机性强,影响因素复杂,传统单一预测模型精度低的问题,结合卷积神经网络(Convolutional neural network,CNN)、双向门控循环单元(Bi-directional gated recurrent unit,BiGRU)以及注意力机制(Attention)在短期电力负荷预测上的不同优点,提出一种基于CNN-BiGRU-Attention的混合预测模型。该方法首先通过CNN对历史负荷和气象数据进行初步特征提取,然后利用BiGRU进一步挖掘特征数据间时序关联,再引入注意力机制,对BiGRU输出状态给与不同权重,强化关键特征,最后完成负荷预测。试验结果表明,该模型的平均绝对百分比误差(Mean absolute percentage error,MAPE)、均方根误差(Root mean square error,RMSE)、判定系数(R-square,R~2)分别为0.167%、0.057%、0.993,三项指标明显优于其他模型,具有更高的预测精度和稳定性,验证了模型在短期负荷预测中的优势。

基于Self-Attention-BiLSTM网络的西瓜种苗叶片氮磷钾含量高光谱检测方法

元素含量无损检测技术可以为植物生长发育的环境精准调控提供关键实时数据。以西瓜苗为例,提出了一种基于图谱特征融合的氮磷钾含量深度学习检测方法。首先,使用高光谱仪拍摄西瓜苗叶片的高光谱图像,使用连续流动化学分析仪测定叶片的3种元素含量。然后,采用基线偏移校正(BOC)叠加高斯平滑滤波(GF)的光谱预处理方法和随机森林算法(RF)建立预测模型,基于竞争性自适应重加权采样(CARS)和连续投影算法(SPA)2种算法初步筛选出特征波长,再综合考虑波长数和建模精度设计了一种最优波长评价方法,将波长数进一步减少到3~4个。最后,提取使用U-Net网络分割的彩色图像颜色和纹理特征,和光谱反射率特征一起作为输入,基于自注意力机制-双向长短时记忆(Self-Attention-BiLSTM)网络构建了3种元素含量的预测模型。实验结果表明,氮磷钾含量预测的R2分别为0.961、0.954、0.958,RMSE分别为0.294%、0.262%、0.196%,实现了很好的建模效果。使用该模型对另2个品种西瓜进行测试,R2超过0.899、RMSE小于0.498%,表明该模型具有很好的泛化性。该高光谱建模方法使用少量波长光谱即实现了高精度检测,在精度和效率上达成了很好的平衡,为后续便携式高光谱检测装备开发奠定了理论基础。

基于CNN-BiLSTM-Attention的三峡库区滑坡地表位移预测研究

地表位移预测在滑坡监测预警中具有重要意义,建立稳定可靠的滑坡位移预测模型是关键。本文基于卷积神经网络和注意力机制的滑坡位移预测方法,并以三峡库区黄泥巴蹬坎滑坡为例进行了验证。本文综合分析了该滑坡长达8年的降雨量、库水位和地表位移等监测数据,建立了结合卷积神经网络(convolutional neural network,CNN)、双向长短期记忆(bidirectional long short-term memory,BiLSTM)网络和注意力机制(attention)的CNN-BiLSTM-Attention深度学习组合预测模型,采用了适应性学习率和正则化技术进行模型训练,提高了模型的泛化能力同时避免过拟合,并与传统LSTM模型进行对比验证。结果表明:相较于传统的机器学习和神经网络方法,该模型在滑坡位移预测精度上取得了显著提升,预测模型拟合优度(R^(2))达0.989,平均绝对百分比误差(MAPE)仅为0.059。

基于特征分解与Bi-LSTM-Attention模型的风向预测

为便于精准控制风电机组的偏航角度、充分利用风能提高机组发电量,提出一种基于历史数据深度学习的风向超短期预测方法。首先利用变分模态分解将风向数据分解成多个子序列,考虑分解后的残差分量仍保留大量信号特征,进一步采用自适应噪声完备集合经验模态分解方法对残差分量进行二次分解。在此基础上,结合风速、环境温度等特征,利用具有注意力机制的双向长短期记忆网络对风向进行超短期预测。采用河北某风电场SCADA真实数据,对风向进行5min的超短期预测实验,并与其他方法进行对比,结果表明所提方法具有更好的风向预测效果。

基于CNN-LSTM-Attention和自回归的混合水位预测模型

【目的】水位预测对交通运输、农业以及防洪措施具有重要影响。精确的水位值可用于提升水道运输的安全及效率、降低洪水风险,同时也是保障区域可持续发展的必要条件。【方法】提出一种CRANet的混合水位预测模型,以卷积神经网络(CNN)、长短期记忆网络(LSTM)、注意力机制以及自回归(AR)组件为基础,旨在应对时间序列数据中存在的线性与非线性问题,缓解自回归及ARIMA模型的缺陷。其应用不仅在于为航运调度提供决策支撑,加强导航安全效率,同样能提升防洪减灾的能力。其中,CNN和LSTM组件有效地针对数据集内的局部和全局关系进行捕捉,AR组件则能充分考虑数据的时间序列特性。同时,通过注意力机制,模型能够优先考虑相关特性,提高预测效果。【结果】研究成果所提出的模型已成功应用于中国西江梧州站的水位预测,在测试集上预测未来3 h级别水位的MAE、RMSE和R^(2)分别为0.086、0.114 5和0.950 8。【结论】结果表明所提出的CRANet模型在水位预测方面的高可用性、准确度与稳健性,相较于AR、SVR、CNN、LSTM等模型具有更优的MAE、RMSE和R^(2)。

基于Coordinate Attention和空洞卷积的异物识别

在我国工厂的工业化生产中,带式运输机占有重要的地位,但是在其运输物料的过程中,常有木板、金属管、大型金属片等混入物料中,从而对带式运输机的传送带造成损毁,引起巨大的经济损失.为了检测出传送带上的不规则异物,设计了一种新的异物检测方法.针对传统异物检测方法中存在的对于图像特征提取能力不足以及网络感受野相对较小的问题,我们提出了一种基于coordinate attention和空洞卷积的单阶段异物识别方法.首先,网络利用coordinate attention机制,使网络更加关注图像的空间信息,并对图像中的重要特征进行了增强,增强了网络的性能;其次,在网络提取多尺度特征的部分,将原网络的静态卷积变为空洞卷积,有效减少了常规卷积造成的信息损失;除此之外,我们还使用了新的损失函数,进一步提高了网络的性能.实验结果证明,我们提出的网络能有效识别出传送带上的异物,较好地完成异物检测任务.

融合MacBERT和Talking⁃Heads Attention实体关系联合抽取模型

针对现有的医学文本关系抽取任务模型在训练过程中存在语义理解能力不足,可能导致关系抽取的效果不尽人意的问题,文中提出一种融合MacBERT和Talking⁃Heads Attention的实体关系联合抽取模型。该模型首先利用MacBERT语言模型来获取动态字向量表达,MacBERT作为改进的BERT模型,能够减少预训练和微调阶段之间的差异,从而提高模型的泛化能力;然后,将这些动态字向量表达输入到双向门控循环单元(BiGRU)中,以便提取文本的上下文特征。BiGRU是一种改进的循环神经网络(RNN),具有更好的长期依赖捕获能力。在获取文本上下文特征之后,使用Talking⁃Heads Attention来获取全局特征。Talking⁃Heads Attention是一种自注意力机制,可以捕获文本中不同位置之间的关系,从而提高关系抽取的准确性。实验结果表明,与实体关系联合抽取模型GRTE相比,该模型F1值提升1%,precision值提升0.4%,recall值提升1.5%。

融合RoBERTa-GCN-Attention的隐喻识别与情感分类模型

在隐喻识别与隐喻情感分类任务的联合研究中,现有多任务学习模型存在对隐喻语料中的上下文语义信息和句法结构信息提取不够准确,并且缺乏对粗细两种粒度信息同时捕捉的问题.针对第1个问题,首先改进了传统的RoBERTa模型,在原有的自注意力机制中引入上下文信息,以此提取上下文中重要的隐喻语义特征;其次在句法依存树上使用图卷积网络提取隐喻句中的句法结构信息.针对第2个问题,使用双层注意力机制,分别聚焦于单词和句子层面中对隐喻识别和情感分类有贡献的特征信息.在两类任务6个数据集上的对比实验结果表明,该模型相比基线模型性能均有提升.