一种基于单层频域自编码器的结构模态分解方法

1.中山大学 航空航天学院,深圳 518107;2.广东省地震局 地震监测与减灾技术重点实验室,广州 510070

模态识别;信号分解;频域自编码器;结构监测

A Structural Mode Separation Method Based on Single Layer Frequency Domain Autoencoder
ZHU Jiajian1,2,XIANG Hong2,WANG Li1,WU Huadeng2,ZHANG Yi2,LYU Zhongrong1

1.School of Aeronautics and Astronautics , Sun Yat-sen University, Shenzhen 518107, China;2.Key Laboratory of Earthquake Monitoring and Disaster Mitigation Technology,Guangdong Earthquake Agency, Guangzhou 510070, China

Modal identification;Signal decomposition;Frequency domain autoencoder;Structure monitoring

DOI: 10.13512/j.hndz.2026.03.13

备注

结构振动模态是表征工程结构动力特性的关键基础信息,在结构设计、振动控制及结构健康监测等多领域具有重要意义。针对基于结构振动监测数据的模态参数识别问题,提出一种基于频域自编码器的模态分离方法。该方法以结构振动响应的傅里叶幅值谱为输入,基于模态叠加原理,通过单层自编码器构建响应信号的解混与重构过程。考虑到理想模态坐标在频域呈现单峰且带宽紧致的特性,设计了一种基于稀疏度量的损失函数,以引导网络学习具有物理意义的模态坐标。在无监督学习框架下,编码层输出用于提取自振频率,解码层权重则用于重构模态振型。通过六层剪切结构数值模型验证了方法的有效性,并进一步应用于某实际三层基础隔震建筑的现场振动监测数据。结果表明:该方法可从测量数据中准确提取结构自振频率及模态振型信息。同时,该方法架构简单,运算效率高,具备良好的工程实用性与实时监测应用潜力。
Structural vibration modes are fundamental indicators of the dynamic characteristics of engineering structures and play a critical role in structural design, vibration control and structural health monitoring. To ad‐dress the problem of modal parameter identification from structural vibration monitoring data,this paper proposes a modal separation method based on a frequency-domain autoencoder. The method takes the Fourier amplitude spec‐trum of structural vibration responses as input and grounded in the principle of modal superposition,employs a sin‐gle-layer autoencoder to model the demixing and reconstruction of response signals. Recognizing that ideal modal co‐ordinates exhibit a single dominant peak with compact bandwidth in the frequency domain, a sparsity-based loss function is designed to guide the network toward learning physically meaningful modal coordinates. Within an unsu‐pervised learning framework, the encoder outputs are used to extract natural frequencies, while the decoder weights are leveraged to reconstruct mode shapes. The effectiveness of the proposed method is first validated through numerical simulations on a six-story shear building model and further demonstrated using field vibration monitoring data from an actual three-story base-isolated structure. Results show that the method accurately identi‐fies natural frequencies and mode shapes. Owing to its simple architecture, the proposed approach demonstrates strong practical engineering applicability,and promising potential for real-time structural monitoring.
·