Rotating Savitzky–Golay Filter Based on BP Neural Network: Enabling High-Accuracy Temperature and Pressure Retrieval via Rayleigh–Brillouin Spectrum Processing

With the development of lidar technology used for spectral detection, the methods corresponding to spectral signal processing have also become a research hot spot. An important issue in spectral signal processing is noise reduction. Commonly, noise reduction methods include the Butterworth low-pass filter, the Wiener filter and the iterative three-point zero-order Savitzky–Golay filter (ITZSGF). However, these methods all have limitations when dealing with spectral noise from Rayleigh-Brillouin lidar. For instance, the low frequency noise below cutoff frequency cannot be removed by Butterworth filter. Additionally, the setting of the singular value threshold in the Wiener filter is complex, and its performance is suboptimal under conditions of low signal-to-noise (SNR). Moreover, the ITZSGF fails to completely remove noise at the both ends of the spectrum. Last but not least, none of the above three methods can eliminate the interference of Mie noise. In order to address these issues, a rotating S-G filter based on BP neural network has been proposed in this work. This method can effectively and simply resolve the problems associated with low frequency noise and noise at the ends of the spectrum, and it can eliminate the influence of Mie noise. Experimental results from air scattering show that the maximum temperature and pressure retrieval errors using the novel method are 1.21K and 0.074bar respectively, which are lower than those obtained using the Butterworth filter, Wiener filter and ITZSGF.