基于深度降噪自编码神经网络的近红外光谱去噪

Denoising of the Near Infrared Spectral Based on Deep Denoising Autoencoder Neural Network

近红外光谱仪在数据采集时,由于受到多种因素的影响,光谱数据常常被一系列噪声所污染,对光谱建模与分析产生巨大的影响。在建模前必须要对数据进行预处理,本文提出一种基于栈式降噪自编码神经网络的光谱信号去噪方法,基于降噪自编码模型重构的思想来实现特征的自动提取,使用无监督逐层贪婪预训练和有监督微调的方法对深度自编码神经网络进行训练,对光谱信号具有良好的噪声滤除效果。与目前比较流行的小波去噪等相比,栈式降噪自编码神经网络有较好去噪性能。最后,本文基于实际光谱仪采集数据进行实验,实验结果验证了该方法的有效性。

In the data acquisition of near infrared spectrometer,due to the influence of many factors,the spectral data are often polluted by a series of noise,which has a great impact on spectral modeling and analysis.Data must be preprocessed before modeling.In this paper,a spectral signal denoising method based on stack noise reduction and self-coding neural network is proposed.Based on the idea of noise reduction and self-coding model reconstruction,the automatic feature extraction is realized.The depth self-coding neural network is trained by using the method of unsupervised layer-by-layer greedy pre-training and supervised fine-tuning,which has a good noise filtering effect on the spectral signal.Compared with the popular wavelet denoising,stack denoising self-coding neural network has better denoising performance.Finally,experiments are carried out based on the actual spectrometer data,and the experimental results verify the effectiveness of the method.