Research On Feature Extraction Method For Acoustic Signal Generation

In the 21st century,the ocean has gradually developed into one of the most critical considerations in the history of a country,and it is closely related to the country’s actual interests.Therefore,our country urgently needs to obtain the priority of ocean development.Whoever seizes the first opportunity in ocean development will win the initiative in future ocean development.Combining the characteristics of the times,the recognition and detection of underwater sound targets,and the generation of underwater sound signals based on sound characteristics are currently one of the important ways for my country to develop marine resources and maintain my country’s marine field.These works are inseparable from the feature extraction of the radiated noise of underwater targets.In recent years,convolutional neural networks have achieved good results in the feature extraction of underwater target radiated noise.Based on the above background,this paper combines convolutional neural networks to propose a feature extraction method for acoustic signal generation.First of all,it is necessary to express the anti-noise characteristics of the radiated noise of underwater targets,which can also be called preprocessing.The LPCC feature representation method has a small amount of calculation and is easy to implement,but has poor anti-noise performance,while the MFCC feature representation method has good recognition performance and anti-noise ability,but requires high calculation volume and accuracy and has no effect on high-frequency processing.Good factors.This paper proposes a convolution weighted feature combination representation method.The underwater target radiated noise is divided into two parts: low frequency and high frequency.The low frequency part is represented by the MFCC feature representation method,and the high frequency part is represented by the LPCC feature representation method.Then the two extracted sound feature vectors are mapped using the mapping interpolation algorithm For vectors of the same size,the two vectors are then weighted and combined using a convolution weighted combination method.This method maximizes the use of the target acoustic signal’s own characteristics.At the same time,the target acoustic signal is segmented and processed by different noise auditory models,which maximizes the characteristics of different auditory models while preserving the noise characteristics.This method can be effectively combined with the subsequent convolutional neural network to achieve the purpose of extracting high-quality features.Secondly,the convolutional neural network has a good tolerance for translation during training,but its tolerance has dropped a lot compared to rotation,and when the fully connected layer integrates the features,the location information and pool of the features will be lost.The transformation operation will abandon the details of the information to a certain extent.A feature extraction method based on a feature enhancement model is proposed.This method is aimed at the feature extraction of underwater target radiation noise.Starting from the convolutional neural network structure,the trained feature set is subjected to affine transformation to form an enhanced feature set.The feature set of and combined with its location features are retrained to obtain more high-quality features.Finally,use the extracted features for sound generation.The experimental results show that the sound generated by the features extracted by this method is more similar to the original sound than the features extracted by the traditional convolutional neural network.