Selective tensorized hybrid Bilstm model for water quality precursor detection

Precursors in multivariate time series (MVTS) data play a critical role in predicting significant events, especially in domains like water quality monitoring. This research proposes a novel Hybrid Selective Tensorized BiLSTM Precursor Detection (HST-BiLSTM-PD) model, which structurally integrates the outputs of two specialized architectures-Selective Tensorized BiLSTM Encoder-Decoder (ST-BiLSTM-ED) and AutoEncoder (ST-BiLSTM-AE)-both enhanced with multi-headed attention mechanisms. The framework consists of two core components: (i) the BiLSTM-AMIL model for bag-level classification of MVTS data into precursor or non-precursor labels, and (ii) the hybrid HST-BiLSTM-PD model, which refines precursor detection by combining the prediction output of the supervised ST-BiLSTM-ED with the reconstruction output of the unsupervised ST-BiLSTM-AE. Furthermore, a per-bag feature-wise relative Mean Square Error thresholding is implemented to enhance the robustness and precision of precursor detection. The model is evaluated on real-world datasets, including the GECCO Industrial Challenge 2019 dataset, Server Machine Dataset (SMD), and DAMADICS dataset, demonstrating its efficacy in capturing intricate temporal patterns and detecting precursors with high accuracy. Tested on the GECCO Industrial Challenge 2019 water quality dataset, the proposed HST-BiLSTM-PD model achieved an F1 score of 0.83, demonstrating its robustness in detecting precursors. Furthermore, tests conducted in the SMD and DAMADICS datasets yielded F1 scores of 0.77 and 0.50, respectively. These results emphasize the model’s versatility and effectiveness in diverse industrial environments, making it a promising tool for real-world precursor detection challenges.