Aerial Acoustics Target Detection Technology Based On Expansion Learning

Aerial acoustics target detection technology is an important research direction in the field of signal processing and object detection.Effective detection of aerial acoustics targets is of great significance for realizing object tracking,battlefield situation awareness and ensuring regional security.However,the signal of helicopters,drones and other low-altitude acoustics targets contain a large amount of noise,which adds difficulty to its recognition and location.Moreover,the difficulty in acquiring the audio of aerial targets and the lack of available data also restrict the development of aerial acoustics detection technology to a large extent.Therefore,how to achieve high-precision real-time detection of aerial acoustics targets in complex environments under the condition of lack of target data has become this article’s research direction.Aiming at the requirements of detecting aerial acoustics target under the condition of small samples and low signal-to-noise ratio,we combine the audio feature extraction technology,the steered response power positioning technology and the WaveNet audio generation technology to improve and innovate the detection method of aerial acoustics target,and proposes an expansion learning method to further improve the recognition accuracy of aerial acoustics target.The concrete research contents of this article are as follows:(1)In view of each link of aerial acoustics target detection,this article elaborates the recognition,location and data generation technology of aerial acoustics target based on the analysis of the characteristics of aerial acoustics target.This article introduces the acoustic signal preprocessing methods such as frame and window,filtering and denoising,analyzes the acoustic target recognition technology based on traditional features and neural network automatic feature extraction,studies the relevant principles of the steered response power algorithm,discusses the working mechanism of the audio generation network,which provides theoretical support for the design of the following algorithms.(2)According to the analysis of the audio characteristics of the two helicopter targets,Black Hawk and Wuzhi 10,corresponding to the feature dimension and statistical functions of the traditional feature set,this article designs a recognition model based on spectrogram characteristics of the aerial target.The model is compared with feature sets+BP neural network in the experiment of target existing recognition and target type recognition.At the same time,about aerial acoustics target location,we improved the search mode of SRPPHAT algorithm based on the quad microphone array,and FNE-SRP-PHAT is proposed by incorporating the noise cancellation algorithm.The detection system is formed by deploying the recognition and location algorithm on the hardware platform,and we completed the test experiment of the location algorithm on it.(3)In order to solve the problem of insufficient aerial acoustics target data,this article uses WaveNet to generate target data in multiple environments,and proposes an interpretable expansion learning method based on multi-layer feature space.First,the validity of the generated data is verified by quantifying the multi-layer feature space.Then 11 training sets and 3 test sets are constructed by mixing the generated data and the original data in different proportions,and we used data set verify the effect of expansion learning.Finally,combined with the mapping relationship between common feature space and specific feature space,we explained the model performance under different expansion ratios,and obtained a recognition model with more generalization ability.