With the popularization of smartphones,smart speakers,speech recognition technology,and autonomous driving technology,the demand and importance of sound classification are constantly increasing.Sound classification refers to the process of categorizing different sound samples into different classes,such as human voice,traffic noise,animal sounds,music,etc.Sound classification has a wide range of applications in many fields,such as security monitoring,smart homes,healthcare,etc.However,one of the main challenges in sound classification is noise interference.In the real world,sound samples are often interfered with by various noise sources,such as wind noise,machine noise,human vo ices,etc.These noises can severely disrupt the accuracy and robustness of sound classification.Therefore,classifying noise types is an important direction in sound classification research.This thesis comprehensively studies the application of self-supervised learning in noise classification,focusing on single and multiple noise source sound classification.The paper analyzes the relevant theories of self-supervised learning,deep learning algorithms,and sound signal processing,and examines the application of self-supervised learning in sound classification and various deep learning algorithms,including convolutional neural networks and residual networks.Finally,through experimental comparison,the effectiveness of these algorithms is evaluated.This thesis proposes two classification algorithms for single and complex noise source sound based on self-supervised learning.For a single noise source,the paper uses the Res Net network combined with self-supervised learning method and Mel feature extraction to design a SNIC(Sim CLR and Res Net-based Mel feature noise classifier)model to classify and recognize noise audio data.By setting model architecture,loss function,and training optimization method,the model is evaluated on various single noise source datasets and achieves 97.85% classification accuracy on the public sound dataset Urban Sound8 k,demonstrating its effectiveness in sound classification.For complex noise sources,the paper proposes a self-supervised learning-based complex noise classification algorithm design,namely CSNIC(Complicated SNIC)algorithm,and provides a detailed analysis of the CSNIC model architecture,loss function design,model training,and optimization.Then,the new complex noise source dataset UE-100 is synthesized based on the public sound datasets Urban Sound8 k and ESC-50,and 84.17% classification accuracy is achieved on the UE-100 dataset,demonstrating its effectiveness in complex noise classification.The two classification algorithms for single and complex noise sources proposed in this thesis exhibit good performance and potential impact in sound classification,providing a way for further exploration and improvement of self-supervised learning in this field. |