Abstract: The multivariate time series data generated during elevator operation are crucial for building an elevator safety monitoring system. To address the issue that existing methods ignore the correlations between sequences, an elevator multivariate time series anomaly detection method based on Dynamic Spatial-Temporal Encoder (DSTE) was proposed. First, a dynamic graph aggregation module was introduced, which computed global node self-attention through graph convolution and attention mechanisms for spatial feature aggregation. Meanwhile, a graph pooling method was employed to dynamically update the graph structure and adjust the degree of correlation between sequences. Then, a temporal feature aggregation module was constructed by incorporating gated convolution modules into traditional temporal convolutional networks and utilizing neural ordinary differential equations for temporal feature propagation, enabling precise spatial-temporal feature modeling. Finally, multilayer convolutions were used for decoding and prediction to identify anomalies. Experimental results on a real-world elevator operation dataset showed that, compared to the suboptimal model DTAAD (Dual TCN-Attention networks for Anomaly Detection), DSTE achieved a 0.35% improvement in precision, a 10.8% increase in recall, a 5.4% enhancement in F1-Score, and a 56.8% reduction in false negative rate, while balancing the false positive rate and false negative rate. The results demonstrate that the proposed method can effectively identify elevator anomalies and reduce the missed detection of abnormal events.

Key words: elevator, multivariate time series, anomaly detection, dynamic graph aggregation, neural ordinary differential equation

摘要： 摘 要: 电梯在运行过程中所产生的多元时间序列数据对于构建电梯安全监测系统至关重要，针对现有方法忽略了序列之间相关性影响的问题，提出一种基于动态时空编码器的电梯多元时间序列异常检测方法（DSTE）。首先，提出动态图聚合模块，通过图卷积与注意力机制计算全局节点自注意力，并进行空间特征聚合；同时采用图池化方法进行动态图结构更新，动态调整序列间的关联程度；其次，构建时域特征聚合模块，向传统时域卷积神经网络添加门控卷积模块，并通过神经常微分方程进行时间特征传播计算，从而实现精准的时空特征建模；最后结合多层卷积进行解码预测，并判断异常。在真实电梯运行数据集中的实验结果表明，与次优模型DTAAD（Dual TCN-Attention networks for Anomaly Detection）相比，DSTE进行异常检测的精确度提升0.35%，召回率提升10.8%，F1Score提升5.4%，漏报率降低56.8%，同时实现了漏报率与误报率之间的平衡，证明了所提方法能够有效识别电梯异常，并减小异常事件漏检率。

关键词: 电梯, 多元时间序列, 异常检测, 动态图聚合, 神经常微分方程