【摘要】 现有基于对比学习的时序异常检测方法大多存在样本对构造复杂，模型资源成本较高的问题。为此提出融合重构与双视角对比学习的时序异常检测方法。对原始时间序列进行频域滤波与时域卷积，从时域和频域两视角生成样本对；使用变分自编码器对原始输入进行重构，通过重构损失与对比损失联合引导编码器部分参数更新；根据编码器的输出的对比损失直接进行异常检测。实验结果表明，所提出方法在较高异常检测效果的同时降低了资源消耗，验证了所设计方法的优势。更多还原

【Abstract】 Most existing contrastive learning-based methods for time series anomaly detection suffer from complex sample pair construction and high resource costs. To address this issue, a method named time series anomaly detection via integrated reconstruction and dual-view contrastive learning(IRDVCL) was proposed. In this method, the original time series was processed by frequency-domain filtering and time-domain convolution to generate sample pairs from both temporal and spectral views. A variational autoencoder was employed to reconstruct the original input, and the encoder parameters were updated under the joint optimization of reconstruction loss and contrastive loss. Anomalies were detected directly based on the contrastive loss of the encoder outputs. Experimental results show that IRDVCL achieves high anomaly detection performance while reducing resource consumption, validating the advantages of the designed method.更多还原