The research of optical fiber Brillouin spectrum denoising based on wavelet transform and neural network

Abstract

The center frequency of Brillouin scattering spectrum is easily influenced by the noise and the measurement accuracy of optical fiber strain is reduced. So a novel denoising method which can be applied in the Brillouin scattering spectrum is developed in this article. The Brillouin scattering spectrum is decomposed into multi-scale detail coefficients and approximation coefficients by using the wavelet transform. The wavelet decomposition detail coefficients are threshold quantified by utilizing the threshold algorithm. At the same time, the wavelet decomposition approximation coefficients are trained and simulated by using the BP neural network in order to remove noise hided in the approximation coefficients. So the novel method can reduce the wavelet decomposition scales. The Brillouin scattering spectrum which has a better denoising effect can be gained by using the inverse wavelet transform, and the measurement accuracy of optical fiber strain is enhanced also. The results of simulation and experiment demonstrate that the proposed method can suppress noise better; accordingly, the new method can gain more precision optical fiber strain and reduce the wavelet decomposition scales effectively than the conventional wavelet denoising method. Theory analysis and experiment show that the method is reasonable and efficient.

Intensifying the response of distributed optical fibre sensors using 2D and 3D image restoration

“…To the best of our knowledge, this is the first time that the high levels of correlation and redundancy contained in the multidimensional domain of the measurements obtained by distributed fibre sensors are exploited for performance improvement. Compared with state-of-the-art methods, the multidimensional processing approach here proposed turns much more efficient than applying known (1D) denoising algorithms 44 45 46 47 48 49 50 51 simply replicated in the different dimensions of interest. For instance, the independent use of 1D processing to denoise time-domain traces and then applied to the processed data in frequency domain leads to denoised data points that do not benefit from the similitude and redundancy that can only be found in a 2D or 3D data structure containing the entire measured information.…”

“…For instance, the independent use of 1D processing to denoise time-domain traces and then applied to the processed data in frequency domain leads to denoised data points that do not benefit from the similitude and redundancy that can only be found in a 2D or 3D data structure containing the entire measured information. For this reason, the method here proposed offers exceptional denoising capabilities when compared with state-of-the-art techniques 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 , enabling a remarkable and unprecedented SNR enhancement, boosting the sensor performance 24 up to about two orders of magnitude with no loss of relevant information at minor added cost. This translates, for instance, into an unmatched 100-fold improvement in the measurand accuracy of conventional distributed sensors.…”

“…The clear understanding reached in recent years on the factors ultimately limiting the sensor capabilities has motivated researchers to propose and demonstrate specially designed methods for performance enhancement 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 . Among several advanced techniques, methods such as distributed Raman amplification 34 35 36 37 , optical pulse coding 38 39 40 41 42 43 or different signal processing methods 44 45 46 47 48 49 50 51 have resulted in implementations outperforming classical standard configurations. Whereas each of these methods can individually provide up to about 10–12 dB SNR enhancement, a higher improvement can only be reached by a proper combination of several of those techniques 52 53 54 , at the cost of complex and expensive implementations.…”

“…Among several existing methods, signal processing techniques, such as optical pulse coding 38 39 40 41 42 43 , wavelet transform 44 45 46 47 48 and Fourier transform 49 have demonstrated to be very efficient tools to remove noise. However, their exploitation for distributed sensing has been restricted so far only to one-dimensional (1D) arrays of data.…”

“…Those techniques can be readily applied, for instance, to Raman-distributed fibre sensors 38 44 , owing to the 1D nature of the acquired data (corresponding to 1D traces of the anti-Stokes, Rayleigh and Stokes backscattered light 12 13 14 ). In the case of Brillouin- and Rayleigh-based distributed sensors, in which time and frequency are scanned, signal processing has been used to denoise individual longitudinal traces (that is, at a fixed scanned frequency) independently from each other 39 40 41 45 , the measured local spectrum at each fibre location 47 or the retrieved measurand profile 48 . Although methods such as time–frequency coding 42 43 take advantage of the double scanning (fibre position and pump–probe frequency detuning) required in Brillouin sensing, the provided SNR enhancement is basically given by the ability of the code to reduce noise in a 1D array of data.…”