题    目：Deep learning in non-stationary signal decomposition

主讲人：周锋 副教授

单    位：广东财经大学

时    间：2025年9月26日 15:30

地    点：7号楼7303

摘    要：The decomposition of non-stationary signals is a fundamental yet complex challenge in time–frequency analysis. While traditional methods such as Empirical Mode Decomposition have laid the foundation for signal decomposition, they suffer from limitations including boundary effects, mode mixing, and low robustness to noise. To address these challenges, recent advances have turned to deep learning approaches, culminating in the development of the Iterative Residual Convolutional Neural Network (IRCNN). IRCNN leverages convolutional neural networks, residual structures, and nonlinear activation functions to perform decomposition by learning the local average of signals in an end-to-end manner. Building upon this, an enhanced framework, IRCNN+, integrates advanced deep learning components—such as multi-scale convolution, attention mechanisms, and total variation denoising—to improve feature representation, smoothness, and adaptability. Comprehensive evaluations demonstrate that both IRCNN and IRCNN+ outperform conventional methods in terms of decomposition accuracy, robustness to noise, and handling of mode mixing and boundary effects, offering a promising direction for scalable and effective non-stationary signal decomposition.

简    介：周锋，广东财经大学南岭学者（卓越人才），信息学院副教授。曾任百度和腾讯推荐算法工程师。美国佐治亚理工学院和意大利拉奎拉大学访问学者。 主持国家自然科学青年基金、 教育部人文社科项目、 广东省自然科学基金面上项目等项目。在IEEE TNNLS, Pattern Recognition, Signal Processing, Expert Systems with Applications, Neural Networks 等期刊发表论文近30篇。研究方向为非平稳信号时频分析，金融大数据分析，推荐算法研究。