Research On Real-time Rain Recognition Based On Sound Signal

Rainfall detection and recognition technology has always been one of the research hotspots in the field of meteorology and hydrology.With the development of artificial intelligence technology,the high attention of signal recognition algorithms provides the possibility to realize innovative and effective rainfall recognition.The change of the rain sound signal can retrieval the development and dissipation process of the rain,which is of great significance for the timely warning of disaster prevention and mitigation.In order to identify the rainfall situation,this paper designs a rainfall recognition system based on sound signals.This system adopts a rain sound collector with a shell,and the number of obtained rain sound signals is 1500,including about 500 signals of light rain(0.1-9.9mm),500 signals of moderate rain(10-24.9mm)and 500 signals of heavy rain(25-49.9mm).A micro rain sound database has been established.In order to extract the purer rain sound,the noisy rain sound signal is processed by the blind source separation algorithm to obtain the rain sound signal component.In order to improve the classification accuracy of the classification model,variational modal decomposition(VMD)combined with wavelet packet algorithm is used to denoise the signal to obtain a pure rain sound signal.Aiming at the disadvantages of poor classification of traditional classification models,two classification studies,namely Random Forest Optimization and Bayesian Optimization Extreme Gradient Boosting(XGBoost)algorithm,were carried out on the sound signals under three rain conditions to obtain better classification results.The specific work is as follows:In order to ensure the reliability of the rain sound signal,a rain sound acquisition device based on STM32 is designed.Considering the sound propagation speed in the medium and preventing unnecessary measurement errors caused by the adhesion and accumulation of rainwater and debris,a conical stainless steel body with an inclination angle of 45 ° is designed,and the microphone collection module is arranged inside.The three ADCs built in the microcontroller are used to collect rain sound signals at the same time and in the same environment,and after the mean value fitting,a signal closer to the actual rain sound is obtained to construct a complete and reliable micro rain sound signal database.In order to obtain pure rain sound,a blind source separation method for rain sound signal processing is proposed.One rain sound signal and two environmental sound signals are combined with a random matrix to mix them into the signal to be observed,and the fixed point algorithm(Fast ICA)and principal component analysis(PCA)algorithm are used to conduct blind source separation experiments.Compare the two blind source separation algorithms.The average running time,the signal-to-noise ratio of the separated signal,and the root mean square error judge the separation effect.The experimental results show that Fast ICA unmixing has a faster convergence rate and the PI separation performance index reaches two decimal places,achieving a better separation effect of blind source separation.Aiming at the problem that rain sound signals are easily affected by interference noise,a rain sound denoising algorithm based on VMD and wavelet packet is proposed.Use VMD algorithm to decompose the rain sound signal into a series of variational modal components(VMF)with sparse characteristics,combine the fast Fourier transform(FFT)to analyze the frequency characteristics of each component,divide the low-frequency,intermediate-frequency and high-frequency components.Use wavelet packet to select different threshold criteria to deal with different frequency bands to achieve denoising.The difference of denoising signal-to-noise ratio and the dirrerences root mean square error are used as evaluation indexes to judge the denoising effect.Experimental results show that the denoising signal-to-noise ratio for the three rain conditions is significantly improved,the root mean square error is reduced by two orders of magnitude,and the denoising effect is improved.Aiming at the shortcomings of traditional neural network algorithms in classifying rain sound signals,two rain sound signal classification algorithms based on decision tree theory are proposed.In order to improve the generalization of the classification model,MFCC and first-order difference coefficients are used to extract features from all rain sound data sets.The first rain sound signal classification algorithm based on random forest adjusts the number of its classification trees to find the best model state for handling the three types of rain.In order to verify the effectiveness of the algorithm,comparing traditional neural network and support vector machine classification models,the classification accuracy rate can reach more than80%.The second rain sound signal classification algorithm based on Bayesian optimization XGBoost uses Bayesian optimization ability to optimize the XGBoost classification model.The experimental results show that the classification accuracy rate has exceeded 90%,and the accuracy rate of light rain and heavy rain signals can reach more than 95%.Compared with the random forest algorithm,the classification effect is better.