Research On Moving Source Location And Classification Recognition Method Based On Recurrent Neural Network

Widespread attention has been given to the location and identification of moving noise sources in both military and civil fields.In military fields,it can serve to identify and track enemy targets,thereby strengthening national defense construction.In civil fields,it can be utilized as a roadside acoustic diagnosis system for monitoring vehicle faults during driving.However,the sound source localization in motion faces numerous challenges due to the complex spatial environment of the sound field and the Doppler effect generated during movement.The traditional algorithm,based on the time difference of arrival(TDOA)and multiple signal classification algorithm(MUSIC),divides the signal into a series of small segments that can be considered as stationary signals for processing non-stationary signals.This approach results in significant positioning errors.Additionally,during the process of acoustic event classification,the complexity of the received target signal and classification algorithm may lead to the occurrence of the "dimension disaster" phenomenon,reducing recognition efficiency.This paper focuses on improving the localization and classification recognition methods of noise sources in motion.A comprehensive investigation is conducted through theoretical derivation,computer simulation,and field experiments.The main contributions of this paper are summarized as follows:(1)A novel position estimation method is proposed in this paper to address the localization of moving noise sources.The method focuses on wayside pass-by noise sources and incorporates Doppler Effect Correction.The acoustic signal is captured using a microphone array,and time-frequency analysis is performed on the signal of each channel.The Approximate Greatest Common Divisor(AGCD)algorithm is utilized to extract the instantaneous frequency.Nonlinear least squares fitting is then applied to estimate the distance of the single-channel sound source based on the instantaneous frequency,followed by obtaining the multi-channel distance difference.The real spatial position of the sound source is ultimately determined using the weighted least squares method.Furthermore,the estimated parameters are employed in the time domain resampling method to mitigate the impact of the Doppler effect.This step allows the intrinsic information of the sound source to be revealed,facilitating the task of environment sound classification(ESC).(2)This paper presents a feature dimensionality reduction architecture that addresses the curse of dimensionality phenomenon in ESC.The proposed architecture combines attention and mutual information to achieve effective dimensionality reduction.Specifically,to align with the two-dimensional MFCC(Mel Frequency Cepstral Coefficients)feature matrix,the method separates and reconstructs the feature frames of different samples.Dimensionality reduction is achieved by leveraging decisions based on the information entropy between the feature frames and labels.Moreover,the method incorporates the LSTM(Long Short-Term Memory)model with an attention mechanism to ensure high recognition accuracy following dimensionality reduction.To validate the performance of the proposed method,ten urban acoustic events from the UrbanSound8k(US8K)dataset are selected for experimentation.The trained model is also transferred to the recognition task of moving sound sources.The simulation results demonstrate that by combining the attention mechanism and mutual information,the proposed method achieves a recognition accuracy of 95.16% on the UrbanSound8 k dataset,while maintaining a small parameter scale of only 0.92 M.Through theoretical analysis,computer simulation,and field experiments,this paper thoroughly investigates the proposed algorithm in terms of location accuracy,acoustic event classification effectiveness,classification model size,and other relevant aspects.The simulation and experimental results consistently indicate that the proposed improved method outperforms the traditional approach across these evaluated aspects.