针对语音识别中DBN-DNN训练时间过长的问题,提出了一种DBN-DNN网络的快速训练方法。该方法从减少误差反向传播计算量的角度出发,在更新网络参数时,通过交替变换网络更新层数来实现加速;同时,也设计了逐渐减少网络全局更新频率和逐渐减...针对语音识别中DBN-DNN训练时间过长的问题,提出了一种DBN-DNN网络的快速训练方法。该方法从减少误差反向传播计算量的角度出发,在更新网络参数时,通过交替变换网络更新层数来实现加速;同时,也设计了逐渐减少网络全局更新频率和逐渐减少网络更新层数两种实施策略。这种训练方法可以与多种DNN加速训练算法相结合。实验结果表明,在不影响识别率的前提下,该方法独立使用或与随机数据筛选(stochastic data sweeping,SDS)算法、ASGD算法等DNN加速训练算法相结合,都可以取得较为理想的加速结果。展开更多
Abundant system operation state information is included in the electrical signal of the hydraulic system motor.How to accurately extract and classify the operation information of electrical signal is the key to realiz...Abundant system operation state information is included in the electrical signal of the hydraulic system motor.How to accurately extract and classify the operation information of electrical signal is the key to realize the condition monitoring of hydraulic system.The early fault characteristics of hydraulic gear pump hidden in the motor current signal are weak and difficult to extract by traditional time-frequency analysis.Based on the correlation coefficient and artificial bee colony algorithm(ABC),the parameter optimization of variational mode decomposition(VMD)is realized in this paper.At the same time,the principle of maximum signal correlation coefficient and kurtosis value is adopted to determine the effective intrinsic mode function(IMF).Moreover,the permutation entropy(PE)and root mean square(RMS)of the effective IMF components are input into the deep belief network(DBN-DNN)as high-dimensional feature vectors.The operation state of gear pump is monitored.The results show that the weak characteristics of current signal of gear pump fault are accurately and stably extracted by this method.The running state of gear pump is monitored and the accuracy of gear fault diagnosis is improved.展开更多
文摘针对语音识别中DBN-DNN训练时间过长的问题,提出了一种DBN-DNN网络的快速训练方法。该方法从减少误差反向传播计算量的角度出发,在更新网络参数时,通过交替变换网络更新层数来实现加速;同时,也设计了逐渐减少网络全局更新频率和逐渐减少网络更新层数两种实施策略。这种训练方法可以与多种DNN加速训练算法相结合。实验结果表明,在不影响识别率的前提下,该方法独立使用或与随机数据筛选(stochastic data sweeping,SDS)算法、ASGD算法等DNN加速训练算法相结合,都可以取得较为理想的加速结果。
基金National Natural Science Foundation of China(No.51675399)。
文摘Abundant system operation state information is included in the electrical signal of the hydraulic system motor.How to accurately extract and classify the operation information of electrical signal is the key to realize the condition monitoring of hydraulic system.The early fault characteristics of hydraulic gear pump hidden in the motor current signal are weak and difficult to extract by traditional time-frequency analysis.Based on the correlation coefficient and artificial bee colony algorithm(ABC),the parameter optimization of variational mode decomposition(VMD)is realized in this paper.At the same time,the principle of maximum signal correlation coefficient and kurtosis value is adopted to determine the effective intrinsic mode function(IMF).Moreover,the permutation entropy(PE)and root mean square(RMS)of the effective IMF components are input into the deep belief network(DBN-DNN)as high-dimensional feature vectors.The operation state of gear pump is monitored.The results show that the weak characteristics of current signal of gear pump fault are accurately and stably extracted by this method.The running state of gear pump is monitored and the accuracy of gear fault diagnosis is improved.