Research On Noise Reduction Of Laser Absorption Spectral Signal Based On Multiple Digital Filtering Techniques And Neural Network

Gas detection technology plays an important role in preventing air pollution.As a new optical detection technology,tunable semiconductor Laser absorption spectroscopy（TDLAS）has the advantages of non-contact,high precision and on-line continuous monitoring compared with traditional technology.However,TDLAS will be interfered by noise from background environment,laser,detector and"etalon effect"in the measurement process.This subject mainly deals with the noise reduction of low concentration NH₃ harmonic signal measured by TDLAS system.Firstly,based on wavelet transform and empirical mode decomposition（EMD）,wavelet-EMD algorithm and cascaded local projection algorithm are used to de-noise the second harmonic signal.The wavelet-EMD and cascaded local projection were used to denoise different noise signals,and the de-noising results were compared with those of improved SVD,wavelet transform and EMD.The results show that cascaded local projection has the best denoising effect for different dB Gaussian white noise reduction among four methods,For optical fringe processing,EMD is the best algorithm,and the noise reduction effect of cascaded local projection is second only to EMD.Considering that the noise in the actual measurement process is of various types,cascaded local projection has the best comprehensive noise reduction effect From the comprehensive treatment of the two kinds of noise.Secondly,In this subject,BP neural network and recurrent neural network（RNN）model are constructed through the Tensorflow platform to denoise the second harmonic signal containing different dB Gaussian white noise.Second harmonic signal with noise is taken as sample input,ideal second harmonic signal is taken as sample label.The trained model is used to process the signal,and the noise reduction results are compared with the traditional digital filtering method.The results show that the denoising effect of neural network is much better than that of traditional digital filtering method when the number of samples is 2000.Among them,RNN has the best denoising effect in all methods and its signal-to-noise ratio is 57dB.For the signal containing 20dB Gaussian white noise,the signal SNR after CAENN denoising reaches 75.10dB and the root mean square error is 0.00003.Finally,In this subject,digital filtering methods such as wavelet-EMD and cascaded local projection and neural network denoising methods such as LSTM and convolutional autocoding are respectively used to de-noise low concentration NH₃ signals collected by TDLAS experiment.and the results are analyzed by using three indicators:signal-to-noise ratio,concentration linear fitting degree and denoising time.The results show that cascaded local projection has the highest signal-to-noise ratio and linear fitting degree of concentration among the all methods,which provides help for accurate measurement of TDLAS.the SNR and concentration linear fitting degree of CAENN are close to cascaded local projection.For the new neural network denoising technique,its significant advantage is that the denoising time is much less than the traditional denoising method,which provides a good choice for TDLAS rapid gas concentration measurement.