桥梁结构健康监测系统采集的加速度数据为结构异常识别和早期损伤检测提供了关键信息。但原始加速度信号复杂，且有较多噪声干扰，这大大降低了结构异常特征可分性，进而增加误判风险。为此，研究提出一种融合多频分解与多尺度时序建模的自监督异常识别框架和一种新的异常敏感指标：高斯核加权残差平方和（GKWRSS）。该方法首先采用经验模态分解（EMD）提取频率敏感分量，随后通过Transformer建立信号的时序长程依赖关系，并对输入信号进行重构，最后运用高斯混合模型（GMM）对输入信号和重构信号的GKWRSS值进行聚类来实现桥梁结构的异常识别。研究结果表明：提出的EMD-Transformer-GMM架构能够有效抑制信号跨频干扰，提升异常特征的可分性，此外，由于Transformer层中引入的转置卷积解码器，架构能够实现自监督学习，在重构输入信号的同时保留细粒度时序特征与频域特异性，同时，该方法仅需基于信号差异进行分类，无需标注异常样本。相较于传统残差平方和（RSS），GKWRSS在保持局部异常敏感性的同时，显著提升了正常与异常状态的可分性。实桥数据和数值模拟试验验证，表明了其在桥梁结构异常识别中的可靠性和实用性。

Acceleration data acquired by bridge structural health monitoring systems provide crucial information for structural anomaly identification and early damage detection. However, the inherent complexity and significant noise contamination of raw acceleration signals substantially reduce the separability of anomalous features, consequently increasing the risk of misjudgment. To address this, a self-supervised anomaly identification framework that integrates multi-frequency decomposition and multi-scale temporal modeling is proposed, along with a novel anomaly-sensitive metric termed the Gaussian Kernel Weighted Residual Sum of Squares (GKWRSS). The proposed method initially employs Empirical Mode Decomposition (EMD) to extract frequency-sensitive components. Subsequently, a Transformer network is utilized to establish long-range temporal dependencies within the signal and reconstruct the input. Finally, anomaly identification is realized by utilizing a Gaussian Mixture Model (GMM) to cluster the GKWRSS values of the original input signal and its reconstructed counterpart. The proposed EMD-Transformer-GMM framework effectively mitigates cross-frequency interference in the signals, thereby enhancing the separability of anomalous features. Furthermore, owing to the transposed convolutional decoder introduced in the Transformer layer, the architecture enables self-supervised learning. It reconstructs the input signal while preserving fine-grained temporal features and frequency-domain specificity. Additionally, this method classifies anomalies based solely on signal discrepancies, eliminating the need for labeled anomaly samples. Compared to the traditional Residual Sum of Squares (RSS), the GKWRSS significantly improves the separability between normal and anomalous states while maintaining sensitivity to local anomalies. Validation using real-world bridge monitoring data demonstrates that the EMD-Transformer-GMM model achieves an average anomaly identification accuracy of 95.73%, confirming its reliability and practical utility for bridge SHM and anomaly detection.