Scholarlink

<p>Real-time, efficient, and accurate air quality anomaly detection is of significant practical importance for traffic navigation and smart cities. Traditional unsupervised deep autoencoders (AE) often struggle to capture the spatiotemporal heterogeneity of air quality anomaly data. Anomaly reconstruction errors are frequently underestimated, and relying solely on one loss function can limit model accuracy and generalizability. It is challenging to perform interpretable spatiotemporal anomaly tracing based solely on model evaluation metrics. This paper proposes a novel air quality anomaly detection and interpretable spatiotemporal traceability analysis method. In the data preprocessing phase, Density Peak Clustering (DPC) combined with Anomaly Score (AS) calculation is utilized to accurately approach the spatiotemporal heterogeneous distribution of various air quality anomalies. Subsequently, we introduce a stochastic derivative-free trust-region algorithm that integrates multiple loss functions, creating the Stochastic derivative-free multi-error-optimized Performer Autoencoder (StoDEMO-PAE). This model replaces the traditional AE structure with a PAE adept at capturing outlier information and optimizes both the loss functions and the data training process. Finally, spatiotemporal traceability analysis is conducted using real-time air quality monitoring data and public complaint information from 95 monitoring stations in Haikou from May 2021 to March 2023. Experimental results demonstrate that StoDEMO-PAE achieves superior average performance on the experimental dataset (P value, R value, and F1 score of 0.788, 0.692, 0.743, respectively, representing an average improvement of 0.175, 0.245, 0.221 over 9 baseline models such as BeatGan, Anomaly Transformer, OmniAnomaly, and LSTM-AE). Furthermore, the integration of public complaint corpus information for air quality anomaly spatiotemporal traceability further validates the scientific robustness and practicality of the proposed method. The results of this study improve air quality anomaly detection, offering valuable insights for real-time urban air quality monitoring, early warnings, and interpretable spatiotemporal traceability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

StoDEMO-PAE: A stochastic derivative-free multi-error-optimized performer autoencoder for air quality anomaly detection and explainable spatiotemporal tracing

  • 期刊论文

  • 发表日期：

    2025年5月9日

查看全文
  • Xiliang Liu,
  • Xiaoying Zhi,
  • Jiashuo Luo,
  • Zhixiang He,
  • Qiang Mei,
  • Tao Zhou,
  • Muyang Liu

摘要

Real-time, efficient, and accurate air quality anomaly detection is of significant practical importance for traffic navigation and smart cities. Traditional unsupervised deep autoencoders (AE) often struggle to capture the spatiotemporal heterogeneity of air quality anomaly data. Anomaly reconstruction errors are frequently underestimated, and relying solely on one loss function can limit model accuracy and generalizability. It is challenging to perform interpretable spatiotemporal anomaly tracing based solely on model evaluation metrics. This paper proposes a novel air quality anomaly detection and interpretable spatiotemporal traceability analysis method. In the data preprocessing phase, Density Peak Clustering (DPC) combined with Anomaly Score (AS) calculation is utilized to accurately approach the spatiotemporal heterogeneous distribution of various air quality anomalies. Subsequently, we introduce a stochastic derivative-free trust-region algorithm that integrates multiple loss functions, creating the Stochastic derivative-free multi-error-optimized Performer Autoencoder (StoDEMO-PAE). This model replaces the traditional AE structure with a PAE adept at capturing outlier information and optimizes both the loss functions and the data training process. Finally, spatiotemporal traceability analysis is conducted using real-time air quality monitoring data and public complaint information from 95 monitoring stations in Haikou from May 2021 to March 2023. Experimental results demonstrate that StoDEMO-PAE achieves superior average performance on the experimental dataset (P value, R value, and F1 score of 0.788, 0.692, 0.743, respectively, representing an average improvement of 0.175, 0.245, 0.221 over 9 baseline models such as BeatGan, Anomaly Transformer, OmniAnomaly, and LSTM-AE). Furthermore, the integration of public complaint corpus information for air quality anomaly spatiotemporal traceability further validates the scientific robustness and practicality of the proposed method. The results of this study improve air quality anomaly detection, offering valuable insights for real-time urban air quality monitoring, early warnings, and interpretable spatiotemporal traceability.