时空变化注意力机制图神经网络的音频事件分类研究

打开文本图片集

中图分类号：TP18文献标识码：A

文章编号：2096-4706（2025）16-0057-07

Research on Audio Event Classification Based on Graph Neural Network with Spatio-Temporal Variation Attention Mechanism

ZHANG Mohua，LIU Ji （School ofComputerandInformation Enginering，Henan Universityof EconomicsandLaw，Zhengzhou45o046，China）

Abstract：Audio event classification faces challenges in complex scenarios，andthe existing methods strugle to capture temporalrelationshpseffectively.Toaddressthis，thispaperproposesaSpati-TmporalVariationAtentionbasedGraphNeural Network （STVA-GNN）， which models audio-visual segments as sequential graph nodes and leverages a Negative Attention Mechanism to compute spatiotemporal variationfeatures betweenadjacent nodes，enhancing intra-modal andcross-modal dynamic information interactions.Thecore inovations include thata Contextual Information Compensation Module （CICM） capturesspatiotemporalevolutionpaterns，andaCross-Modal Gaph Variation IncentiveModule（CMGVI）enhancesaudionode weightsusingvideo-modalspatiotemporalvariations fordepfusion.ExperimentalresultsontheAudioSetdatasetdemonstrate that STVA-GNNachieves mAPandAUC scores of 0.56and0.94respectively，outperforming mainstream methods.Additionaly it maintains a significant advantage in noisy environments，verifying itsrobustness.

Keywords：audioeventclasification;Spatio-TemporalVariationAtentionMechanism;Temporal GraphNeuralNetwork; change information compensation;cross-modal information fusion

0 引言

音频信号是现实世界信息传递的关键载体，尤其在人工智能领域中，音频事件分析已成为研究的前沿热点。（剩余11893字）