Abstract: Time series analysis provides important data support for scientific decision-making, resource optimization, and the construction of risk prevention systems by accurately capturing the variation patterns of environmental factors. In complex multivariate time series analysis scenarios, the efficient extraction of trend information and the precise modeling of variable relationships are crucial for enhancing the model’s analytical capabilities across all scenarios. To address the limitations of existing Transformer-based models in extracting trend information, as well as the distortion in modeling variable correlations caused by sensor heterogeneity, this paper proposes a multivariate time series data analysis method that integrates a transposed self-attention mechanism with a convolutional autoencoder. The proposed method constructs a transposed one-dimensional convolutional autoencoder structure to achieve feature reconstruction along the temporal dimension, effectively capturing long-term trend information in multivariate time series data. At the same time, the transposed self-attention mechanism is incorporated into the multivariable correlation analysis process, significantly improving the accuracy of modeling inter-variable relationships. Comparative experiments on predictive performance against recent state-of-the-art methods—including iTransformer, TimeMixer, TimesNet, DLinear, and Autoformer—demonstrate that the proposed method reduces the MSE metric by 11.4%, 3.1%, 5.2%, 7.4%, and 32.7%, respectively. Furthermore, anomaly detection experiments conducted on five public datasets show that the average F1 score of the proposed method outperforms that of other baseline approaches. Notably, prediction experiments on a real-world water quality monitoring dataset indicate that the proposed method achieves up to 46.1% and 42.9% reductions in average MAE and MSE metrics, respectively, fully validating its effectiveness in practical application scenarios.

Key words: Anomaly detection, Time series forecasting, Self-attention mechanism, Convolutional autoencoder, Environmental monitoring

摘要： 时序分析通过精准捕捉环境要素的变化规律，为科学决策制定、资源优化配置和风险防控体系构建提供了重要的数据依据。在复杂多元时间序列分析场景中，趋势性信息的高效提取与变量间关系的精准建模，是有效提升模型全场景分析能力的关键所在。针对现有Transformer类模型在趋势性信息提取方面存在的不足，以及传感器异构性导致的变量相关性建模失真等问题，提出了一种融合转置自注意力机制与卷积自编码器的多元时序数据分析方法。该方法通过构建转置一维卷积自编码器结构，从时间维度实现时序数据的特征重建，有效捕获多元时序数据中的长期趋势信息；同时将转置的自注意力机制引入多变量相关性分析过程，显著提升了变量间关联关系的建模精度。在与iTransformer、TimeMixer 、TimesNet、DLinear和Autoformer等近年主流方法的预测性能对比实验中，本方法的MSE指标分别降低了 11.4%、3.1%、5.2%、7.4%和32.7%。此外，在5个公共数据集上开展的异常检测实验表明，该方法的平均F1分数优于其他基线方法。特别在真实水质监测数据集上的预测实验中，与现有方法相比，本方法的平均MAE和MSE指标最高降低了46.1%和42.9%，充分验证了方法在实际应用场景中的有效性。

关键词: 异常检测, 时序预测, 自注意力机制, 卷积自编码器, 环境监测