Anomaly detection based on spatiotemporal LSTM with Transformer-AE generative adversarial networks

As digital intelligent industries continue to evolve, the time series data generated by industrial sensors has become increasingly complex and continues to expand, complicating timely anomaly detection. Therefore, time series anomaly detection has become particularly important. These time series data exhibit complex contextual dependencies in both temporal and spatial dimensions, while also facing challenges related to dynamic feature selection. Additionally, in a high-dimensional data environment, the complexity and cost issues associated with anomaly identification and labeling are becoming more pronounced. To address these issues, this research proposes a novel anomaly detection model— LSTransGAN-AE. The feature extraction module, LSTM-TS, blends spatiotemporal attention mechanisms with Long Short-Term Memory (LSTM) networks, effectively enhancing the feature extraction capability for time series data. The detection module, TransGAN-AE, integrates Generative Adversarial Networks (GAN) with Transformer Autoencoders (AEs) to improve detection precision. To further enhance the model's anomaly detection accuracy, we introduce Gaussian filtering for smoothing, thereby reducing false positive results. Finally, we validate the effectiveness of this method on multiple time series datasets (SWat, SMD, and ECG). Compared to other methods, our model achieved average improvements of 0.123, 0.150, and 0.155 in AUPRC, AUROC, and F1 metrics, respectively. These results indicate that the proposed LSTransGAN-AE model performs exceptionally well in anomaly detection tasks and demonstrates strong application potential.