针对深度学习应用技术进行了研究性综述。详细阐述了RBM(受限玻尔兹曼机)逐层预训练后再用BP(反向传播)微调的深度学习贪婪层训练方法,对比分析了BP算法中三种梯度下降的方式,建议在线学习系统采用随机梯度下降,静态离线学习系统采用随...针对深度学习应用技术进行了研究性综述。详细阐述了RBM(受限玻尔兹曼机)逐层预训练后再用BP(反向传播)微调的深度学习贪婪层训练方法,对比分析了BP算法中三种梯度下降的方式,建议在线学习系统采用随机梯度下降,静态离线学习系统采用随机小批量梯度下降;归纳总结了深度学习深层结构特征,并推荐了目前最受欢迎的五层深度网络结构设计方法。分析了前馈神经网络非线性激活函数的必要性及常用的激活函数优点,并推荐Re LU(rectified linear units)激活函数。最后简要概括了深度卷积神经网络、深度递归神经网络、长短期记忆网络等新型深度网络的特点及应用场景,归纳总结了当前深度学习可能的发展方向。展开更多
Failure depth of coal seam floors is one of the important considerations that must be kept in mind when mining is carried out above a confined aquifer. In order to study the factors that affect the failure depth of co...Failure depth of coal seam floors is one of the important considerations that must be kept in mind when mining is carried out above a confined aquifer. In order to study the factors that affect the failure depth of coal seam floors such as mining depth, coal seam pitch, mining thickness, workface length and faults, we propose a combined artificial neural networks (ANN) prediction model for failure depth of coal seam floors on the basis of existing engineering data by using genetic algorithms to train the ANN. A practical engineering application at the Taoyuan Coal Mine indicates that this method can effectively determine the network struc- ture and training parameters, with the predicted results agreeing with practical measurements. Therefore, this method can be applied to relevant engineering projects with satisfactory results.展开更多
Data Dependent Superimposed Training(DDST) scheme outperforms the traditional su-perimposed training by fully canceling the effects of unknown data in channel estimator.In DDST,however,the channel estimation accuracy ...Data Dependent Superimposed Training(DDST) scheme outperforms the traditional su-perimposed training by fully canceling the effects of unknown data in channel estimator.In DDST,however,the channel estimation accuracy and the data detection or channel equalization performance are affected significantly by the amount of power allocated to data and superimposed training sequence,which is the motivation of this research.In general,for DDST,there is a tradeoff between the channel estimation accuracy and the data detection reliability,i.e.,the more accurate the channel estimation,the more reliable the data detection;on the other hand,the more accurate the channel estimation,the more demanding on the power consumption of training sequence,which in turn leads to the less reliable data detection.In this paper,the relationship between the Signal-to-Noise Ratio(SNR) of the data detector and the training sequence power is analyzed.The optimal power allocation of the training sequence is derived based on the criterion of maximizing SNR of the detector.Analysis and simulation results show that for a fixed transmit power,the SNR and the Symbol Error Rate(SER) of detector vary nonlinearly with the increasing of training sequence power,and there exists an optimal power ratio,which accords with the derived optimal power ratio,among the data and training sequence.展开更多
Electric load forecasting has been a major area of research in the last decade since the production of accurate short-term forecasts for electricity loads has proven to be a key to success for many of the decision mak...Electric load forecasting has been a major area of research in the last decade since the production of accurate short-term forecasts for electricity loads has proven to be a key to success for many of the decision makers in the energy sector, from power generation to operation of the system. The objective of this research is to analyze the capacity of the MLP （multilayer perceptron neural network） versus SOM （self-organizing map neural network） for short-term load forecasting. The MLP is one of the most commonly used networks. It can be used for classification problems, model construction, series forecasting and discrete control. On the other hand, the SOM is a type of artificial neural network that is trained using unsupervised data to produce a low-dimensional, discretized representation of an input space of training samples in a cell map. Historical data of real global load demand were used for the research. Both neural models provide good prediction results, but the results obtained with the SOM maps are markedly better Also the main advantage of SOM maps is that they reach good results as a network unsupervised. It is much easier to train and interpret the results.展开更多
文摘针对深度学习应用技术进行了研究性综述。详细阐述了RBM(受限玻尔兹曼机)逐层预训练后再用BP(反向传播)微调的深度学习贪婪层训练方法,对比分析了BP算法中三种梯度下降的方式,建议在线学习系统采用随机梯度下降,静态离线学习系统采用随机小批量梯度下降;归纳总结了深度学习深层结构特征,并推荐了目前最受欢迎的五层深度网络结构设计方法。分析了前馈神经网络非线性激活函数的必要性及常用的激活函数优点,并推荐Re LU(rectified linear units)激活函数。最后简要概括了深度卷积神经网络、深度递归神经网络、长短期记忆网络等新型深度网络的特点及应用场景,归纳总结了当前深度学习可能的发展方向。
基金Projects 50874103 supported by the National Natural Science Foundation of China2006CB202210 by the National Basic Research Program of China+1 种基金BK2008135 by the Natural Science Foundation of Jiangsu Provincethe Qing-lan Project of Jiangsu Province
文摘Failure depth of coal seam floors is one of the important considerations that must be kept in mind when mining is carried out above a confined aquifer. In order to study the factors that affect the failure depth of coal seam floors such as mining depth, coal seam pitch, mining thickness, workface length and faults, we propose a combined artificial neural networks (ANN) prediction model for failure depth of coal seam floors on the basis of existing engineering data by using genetic algorithms to train the ANN. A practical engineering application at the Taoyuan Coal Mine indicates that this method can effectively determine the network struc- ture and training parameters, with the predicted results agreeing with practical measurements. Therefore, this method can be applied to relevant engineering projects with satisfactory results.
基金the National Natural Science Foundation of China(NSFC)(No.60472089)
文摘Data Dependent Superimposed Training(DDST) scheme outperforms the traditional su-perimposed training by fully canceling the effects of unknown data in channel estimator.In DDST,however,the channel estimation accuracy and the data detection or channel equalization performance are affected significantly by the amount of power allocated to data and superimposed training sequence,which is the motivation of this research.In general,for DDST,there is a tradeoff between the channel estimation accuracy and the data detection reliability,i.e.,the more accurate the channel estimation,the more reliable the data detection;on the other hand,the more accurate the channel estimation,the more demanding on the power consumption of training sequence,which in turn leads to the less reliable data detection.In this paper,the relationship between the Signal-to-Noise Ratio(SNR) of the data detector and the training sequence power is analyzed.The optimal power allocation of the training sequence is derived based on the criterion of maximizing SNR of the detector.Analysis and simulation results show that for a fixed transmit power,the SNR and the Symbol Error Rate(SER) of detector vary nonlinearly with the increasing of training sequence power,and there exists an optimal power ratio,which accords with the derived optimal power ratio,among the data and training sequence.
文摘Electric load forecasting has been a major area of research in the last decade since the production of accurate short-term forecasts for electricity loads has proven to be a key to success for many of the decision makers in the energy sector, from power generation to operation of the system. The objective of this research is to analyze the capacity of the MLP （multilayer perceptron neural network） versus SOM （self-organizing map neural network） for short-term load forecasting. The MLP is one of the most commonly used networks. It can be used for classification problems, model construction, series forecasting and discrete control. On the other hand, the SOM is a type of artificial neural network that is trained using unsupervised data to produce a low-dimensional, discretized representation of an input space of training samples in a cell map. Historical data of real global load demand were used for the research. Both neural models provide good prediction results, but the results obtained with the SOM maps are markedly better Also the main advantage of SOM maps is that they reach good results as a network unsupervised. It is much easier to train and interpret the results.
