Micro Vibration Signal Extraction Technology Based On Deep Neural Network

Having the advantages of high precision,long distance and non-contact,microvibration measurement based on laser detection is widely used in national defense,industrial production and scientific research.Based on reviewing the related research in this field and the theory of vibration measurement using laser speckle,this thesis attempts to extract and rebuild the weak vibration signal of an object by using deep neural network.The method presented in this paper is innovative and can be applied in engineering experience.The results can be a reference for related research in this field.In this paper,a variety of speech audios are used as the excitation source of weak vibration signal,and a 532 nm laser is casted on the surface of the vibration body,so as to produce the laser speckle modulated by the vibration signal.During the experiment,the type of input speech,detection distance,source speech quality and some other experimental parameters are changed,a high-speed linear array CMOS camera is used to record the speckle sequence.The recorded image is converted into a speech signal by transforming,removing DC component,and amplifying.Then,the initial speech signals are used as inputs for various deep neural network pre-trained by cloud speech database separately for enhancement processing.Finally,the commonly used international speech evaluation methods such as segSNR(Segmental Signal-to-Noise Ratio),LLR(Log Likelihood Ratio),NSEC(normalized subband envelope correlation)and PESQ(perceptual evaluation of speech quality)are used to evaluate the enhancement effect.Three types of deep neural network are exploited in the experiment.The first type is convolutional neural network(CNN),which performs well in the field of noise reduction.The second type is gated recurrent unit(GRU),which is a variant of recurrent neural network widely used in time series processing.The third type is variational autoencoder(VAE),a typical noise reduction network of unsupervised learning.The combination effect of VAE and CNN,VAE and GRU is also analyzed and explored.For each group of experiments,ten samples are used as input.The average value of sample results is used as the basis for evaluation.The results show that the three types of deep network can effectively improve the quality of speech signal in reducing noise and improving intelligibility,and the effect of combining the two types of neural network is better than that of using one type of network alone.Among the network,the combination of VAE and CNN gets the generally best result,which can improve 3.6 dB in segSNR,-0.2 in LLR,0.18 in NSEC and 1.5 in PESQ score.The score of combination of VAE and GRU is slightly lower,the improvement degree of the above scores is 3.4 dB,-0.18,0.15 and 1.35.Considering results of using one single neural network,CNN has the best effect,GRU gets the middle,and VAE is slightly lower than the other two.The experimental results show that neural network could effectively enhance the quality of the extracted speech signal.Finally,the method of generating noisy audio by using the noise existing in the experimental data and training convolutional neural network to process audio files is proposed,the experimental results show that the score of segSNR was improved by 3-4db and the score of NSEC was improved by about 0.14,which proves the effectiveness of the proposed scheme.