Robust time-series anomaly detection under structural distortions: integrating Hybrid AWRED v5 into classical and bidirectional LSTM networks

DOI: 10.31673/2412-9070.2026.318104

Abstract

Deep recurrent neural networks are widely used for time-series anomaly detection, but their per-formance strongly depends on the purity of the training data. The classical Mean Squared Error (MSE) loss is sensitive to structural distortions because anomalous windows can bias gradient updates and degrade the quality of the learned latent representation. This paper considers the fifth generation of the Hybrid AWRED v5 methodology (Adaptive Weighted Reconstruction with Regularized Energy and Dynamics), which introduces a differentiable loss function with a spatio-temporal soft-thresholding mechanism. Bayesian optimization is used to tune the weighting function parameters, allowing the clipping threshold to adapt to the current distribution of reconstruction errors.
The experimental study compares two baseline architectures, a classical LSTM and a bidirectional BiLSTM with layer normalization, under a 15% structural contamination rate in the training sample. With the standard MSE loss, the Average Precision metric reaches 0.534 for the classical LSTM and 0.566 for the bidirectional architecture. After integrating Hybrid AWRED v5, these values increase to 0.835 and 0.922, respectively. For the modern architecture, the proposed objective function also achieves an AUC-ROC of 0.990. These results indicate that the method preserves anomaly ranking quality more effectively under contaminated training conditions.
Overall, Hybrid AWRED v5 can be viewed as a promising architecture-agnostic tool for robust optimization in time-series anomaly detection. At the same time, its advantages are most directly supported by detection and ranking metrics, while a separate quantitative evaluation of spatio-temporal localization remains a direction for further work.

Keywords: anomaly detection, time series, deep learning, LSTM, BiLSTM, structural distortions, data contamination, robust optimization, Hybrid AWRED v5.