【作者】 王立红； 张延华； 李梦璐； 贾克斌；

【Author】 WANG Lihong;ZHANG Yanhua;LI Menglu;JIA Kebin;School of Information and Communication Engineering, Beijing University of Technology;

【机构】 北京工业大学信息与通信工程学院；

【摘要】 在具有强噪声的工业环境中,传感器数据可能混有强的现场背景噪声,使得识别工业环境的信号变得困难。强工业噪声严重污损信号原有特征和可辨识度。为去除信号中混入的强噪声,针对U-Net网络,通过新增Dropout层,防止数据的过拟合;并引入带泄露修正单元函数(LeakyReLU),解决梯度消失问题;同时使用随机梯度下降(Stochastic Gradient Descent,SGD)优化算法,结合回归损失函数(Huber Loss)训练模型,降低了损失值。仿真采用混有强工业现场噪声的语音信号,通过间接预测残差噪声谱替代直接预测去噪语音谱,减少了训练时间。实验结果表明,全局残差U-Net模型可以实现对工业强噪声的有效抑制。更多还原

【Abstract】 In an industrial environment with strong noise, sensor data may be mixed with strong on-site background noise, making it difficult to identify signals from the industrial environment. Strong industrial noise severely defaces the original characteristics and recognizability of the signal. In order to remove the strong noise mixed in the signal, the paper aims at the U-Net network, by adding a new Dropout layer to prevent overfitting of the data; and introducing a leak correction unit function(Leaky ReLU) to solve the problem of gradient disappearance. At the same time, the stochastic gradient descent(SGD) optimization algorithm is used, combined with the regression loss function(Huber Loss) to train the model, which reduces the loss value. The simulation uses speech signals mixed with strong industrial site noise, and indirectly predicts the residual noise spectrum instead of directly predicting the denoised speech spectrum, reducing the training time. The experimental results show that the global residual U-Net model can achieve strong industrial noise effective suppression.更多还原