有效波高(significant wave height,SWH)是海洋的重要参数之一,对其的精确预测对渔业发展、海上交通和海洋生态系统具有重要意义。为了提高有效波高的预测精度,本文提出了一种基于卷积神经网络-时空长短时记忆神经网络-卷积神经网络(con...有效波高(significant wave height,SWH)是海洋的重要参数之一,对其的精确预测对渔业发展、海上交通和海洋生态系统具有重要意义。为了提高有效波高的预测精度,本文提出了一种基于卷积神经网络-时空长短时记忆神经网络-卷积神经网络(convolutional neural network-spatiotemporal long short-term memory-convolutional neural network,CNN-STLSTM-CNN)的有效波高预测模型。该模型由编码器(Encoder)、解释器(Translator)和解码器(Decoder)构成。Encoder通过卷积神经网络提取SWH数据的空间特征,Translator通过时空长短时记忆神经网络(spatiotemporallongshort-term memory,STLSTM)提取SWH数据的空间特征在时间上的变化特性,Decoder通过卷积神经网络的转置卷积模块重建预测结果。对东海和南海海域的二维有效波高数据进行建模,实验结果表明CNNSTLSTM-CNN模型的均方根误差(root mean squared error,RMSE)、平均绝对误差(mean absolute error,MAE)、均方根误差均值(mean of root mean squared error,M_RMSE)和平均绝对误差均值(mean of mean absolute error,M_MAE)等指标值均低于已有的方法,验证了CNN-STLSTM-CNN模型的有效性。展开更多
目前对人体姿态骨骼关键点检测存在两个研究难点,一是如何由2D姿态进行3D人体姿态估计,另一个是标准数据库和用户上传的视频动作在时间上不匹配。为此,本文提出基于空洞转置卷积的沙漏结构(Dilated and Transpose Convolutions Hourglas...目前对人体姿态骨骼关键点检测存在两个研究难点,一是如何由2D姿态进行3D人体姿态估计,另一个是标准数据库和用户上传的视频动作在时间上不匹配。为此,本文提出基于空洞转置卷积的沙漏结构(Dilated and Transpose Convolutions Hourglass,DTCH)神经网络;然后应用卡尔曼滤波算法进行数据降噪处理,最后利用动态时间规整(Dynamic Time Warping,DTW)算法提高患者运动时姿态匹配的准确性。在仿真实验中,该模型在Human3.6M数据集上的平均每关节位置误差(MPJPE)与相关研究的最佳结果相比减少了11%,可以精确高效地实现3D人体姿态估计。展开更多
The paper proposed a prediction method of combustion temperature field in a coal-fired boiler of a 350 MW unit through deep learning.The method utilizes operating parameters and multi-point temperature data as inputs ...The paper proposed a prediction method of combustion temperature field in a coal-fired boiler of a 350 MW unit through deep learning.The method utilizes operating parameters and multi-point temperature data as inputs for online predicting temperature field.Firstly,to establish the mapping relationship between temperature field and operating parameters as well as multi-point temperature data,a data set was constructed.In the data set,the temperature fields were obtained through the inversion of thermal radiation imaging model,while the operating parameters were collected from the distributed control system of the unit.Then,a transpose convolutional neural network(TCNN)model was developed to obtain the mapping relationship based on the data set.In the simulation study,multi-point temperature data were obtained through the forward calculation of the thermal radiation imaging model.The impact of the quantity and location of multi-point temperature data on generalization ability of the TCNN model was analyzed.In the experimental study,multi-point temperature data were measured by image probes.A comparative analysis was conducted to evaluate generalization ability of the TCNN model with and without the addition of multi-point temperature data,benchmarking against existing methods.With the addition of multi-point temperature data,the mean absolute percentage errors of predicted temperature fields are all less than 1.6%at four stable loads,while the maximum relative error of average value of predicted temperature field decreases from 7.24%to 2.77%during variable load process.The proposed prediction method has promising potential for online combustion monitoring in the furnace.展开更多
文摘有效波高(significant wave height,SWH)是海洋的重要参数之一,对其的精确预测对渔业发展、海上交通和海洋生态系统具有重要意义。为了提高有效波高的预测精度,本文提出了一种基于卷积神经网络-时空长短时记忆神经网络-卷积神经网络(convolutional neural network-spatiotemporal long short-term memory-convolutional neural network,CNN-STLSTM-CNN)的有效波高预测模型。该模型由编码器(Encoder)、解释器(Translator)和解码器(Decoder)构成。Encoder通过卷积神经网络提取SWH数据的空间特征,Translator通过时空长短时记忆神经网络(spatiotemporallongshort-term memory,STLSTM)提取SWH数据的空间特征在时间上的变化特性,Decoder通过卷积神经网络的转置卷积模块重建预测结果。对东海和南海海域的二维有效波高数据进行建模,实验结果表明CNNSTLSTM-CNN模型的均方根误差(root mean squared error,RMSE)、平均绝对误差(mean absolute error,MAE)、均方根误差均值(mean of root mean squared error,M_RMSE)和平均绝对误差均值(mean of mean absolute error,M_MAE)等指标值均低于已有的方法,验证了CNN-STLSTM-CNN模型的有效性。
文摘目前对人体姿态骨骼关键点检测存在两个研究难点,一是如何由2D姿态进行3D人体姿态估计,另一个是标准数据库和用户上传的视频动作在时间上不匹配。为此,本文提出基于空洞转置卷积的沙漏结构(Dilated and Transpose Convolutions Hourglass,DTCH)神经网络;然后应用卡尔曼滤波算法进行数据降噪处理,最后利用动态时间规整(Dynamic Time Warping,DTW)算法提高患者运动时姿态匹配的准确性。在仿真实验中,该模型在Human3.6M数据集上的平均每关节位置误差(MPJPE)与相关研究的最佳结果相比减少了11%,可以精确高效地实现3D人体姿态估计。
基金supported by the National Key R&D Program of China(2024YFB4104804).
文摘The paper proposed a prediction method of combustion temperature field in a coal-fired boiler of a 350 MW unit through deep learning.The method utilizes operating parameters and multi-point temperature data as inputs for online predicting temperature field.Firstly,to establish the mapping relationship between temperature field and operating parameters as well as multi-point temperature data,a data set was constructed.In the data set,the temperature fields were obtained through the inversion of thermal radiation imaging model,while the operating parameters were collected from the distributed control system of the unit.Then,a transpose convolutional neural network(TCNN)model was developed to obtain the mapping relationship based on the data set.In the simulation study,multi-point temperature data were obtained through the forward calculation of the thermal radiation imaging model.The impact of the quantity and location of multi-point temperature data on generalization ability of the TCNN model was analyzed.In the experimental study,multi-point temperature data were measured by image probes.A comparative analysis was conducted to evaluate generalization ability of the TCNN model with and without the addition of multi-point temperature data,benchmarking against existing methods.With the addition of multi-point temperature data,the mean absolute percentage errors of predicted temperature fields are all less than 1.6%at four stable loads,while the maximum relative error of average value of predicted temperature field decreases from 7.24%to 2.77%during variable load process.The proposed prediction method has promising potential for online combustion monitoring in the furnace.
