Research On Multi-Dimensional Feature Extraction Methods For Ship Radiated Noise

The classification and recognition of ship radiated noise targets is a key research topic in the field of underwater acoustics,and the method of extracting target features is an important support in classification and recognition technology.The extracted target features will affect the effectiveness of subsequent target recognition.Due to the complexity and variability of the marine environment,it is difficult to accurately represent target features with a single feature quantity.Therefore,it is necessary to extract target feature information from multiple dimensions to obtain complete target features and achieve better classification results.This article analyzes the characteristics of ship radiated noise,investigates feature extraction methods in frequency domain,modal domain,and auditory domain,and conducts multi-dimensional feature extraction based on publicly available measured datasets on the network.Finally,the extracted multi-dimensional feature vectors are combined to form a multidimensional feature set and analyze the separability of the feature set.The main content of this paper is:1.Studied the characteristics of ship radiated noise signals.Introduce the state information of the ship itself into the modeling process of ship radiated noise to generate simulation signals that are close to the real ship radiated noise.Three spectral feature analysis methods were elaborated: continuous spectrum,LOFAR spectrum,and DEMON spectrum.And based on the publicly measured dataset of the network,spectral features of different targets were extracted,and it was verified that the feature vectors extracted by the three methods can express the characteristics of different targets,providing a basis for the construction of multidimensional feature sets of subsequent targets.2.Studied empirical mode decomposition methods and noise assisted mode decomposition methods,and compared the decomposition effects of different mode decomposition methods.Due to the different feature information contained in each order of IMF after decomposition of different target modes,the average instantaneous frequency,center frequency,energy density,and energy allocation proportion of the IMF are used as the target mode feature quantities.Extracting the modal features of the measured data IMF,it was found that the distribution of modal feature quantities varies among different targets,and modal feature quantities can be used as the basis for target classification.In addition,a denoising method for signal reconstruction using modal decomposition was studied,and the denoising performance of this method was demonstrated through simulation experiments that calculated the signal similarity and signal-to-noise ratio before and after denoising.3.Studied the feature extraction method for ship radiated noise based on auditory features,and introduced commonly used speech signal processing algorithms.Due to the similarity between speech signals and underwater acoustic signals,the MFCC and GFCC algorithms in speech signals are migrated to underwater acoustic signal processing.The results indicate that the MFCC and GFCC features of different targets exhibit differences with increasing order.A method for extracting auditory cepstrum coefficients based on modal decomposition denoising was studied.Support vector machines were used to identify 356 samples tested,and the results showed that the recognition rate before denoising was 96.43%,while the recognition rate after denoising was 97.71%.4.Build a multidimensional feature set for the Deep Ship dataset publicly available on the internet.Extract the continuous spectrum,line spectrum,DEMON spectrum,modal frequency,modal energy,and auditory features of four types of ship targets in the dataset.Perform separability analysis on the above features using support vector machines and distance based separability criteria,respectively.The results indicate that different targets with single feature recognition have their own advantages,but there is also the problem of similar targets being indistinguishable.Therefore,continuous spectrum,line spectrum,and DEMON spectrum are combined as spectral features,and modal frequency and energy are combined as modal features.Finally,spectral features,modal features,and auditory features are constructed as multidimensional feature sets.The separability analysis results show that the maximum average distance between the feature sets formed by modal features and auditory features in the four combinations is 0.71,and the recognition accuracy is 98.14%.The other three combination methods have a small average spatial distance and a recognition accuracy of less than 65%.