A significant number of tunnels are equipped with health monitoring systems that generate multi-source monitoring series. These series exhibit pronounced mixed periodicity, containing both long-term and short-term patterns. In such circumstances, conventional methodologies, such as autoregressive models and Gaussian processes, may prove ineffective in capturing the intricate temporal dependencies. This paper presents an enhancement to the LSTNet model and introduces BiLSTNet-MHA, a data prediction model designed for the extraction of mixed-period features in tunnel monitoring data. The model facilitates a more comprehensive extraction of mixed-period features and reduces computational complexity.

BiLSTNet-MHA is an extension of the LSTNet architecture. Convolutional layers with a one-dimensional CNN can capture short-term local dependencies in multivariate series. A bidirectional long short-term memory network, in conjunction with an LSTM-skip layer, extracts bidirectional long-term dependencies and periodic patterns in the series. A multi-head attention mechanism is introduced as a self-attention layer to capture non-periodic temporal patterns. A conventional autoregressive model constitutes the linear component, compensating for the limitations of neural networks in processing scale variations. Singular value decomposition is employed to perform low-rank approximation and dimensionality reduction for spatially correlated multivariate series. The training process utilizes quantile loss at 0.25, 0.5, and 0.75. The model is applied to spatially correlated structural-response monitoring series and factor-correlated environmental load series from a domestic tunnel. The structural responses include crown settlement, surrounding rock pressure, and concrete stress. The environmental loads include temperature, humidity, and wind speed. The monitoring period extends from January 1, 2020, to December 31, 2020. The records are resampled to 30-minute intervals, and the dataset is organized by quarter for analysis, with 80% of the data allocated for training and 20% for testing. The implementation utilizes PyTorch in conjunction with Python 3.7.11. The mean absolute error, root mean squared error, and the coefficient of determination are used to evaluate the forecasting performance of each model. Ablation variants omit recurrent skipping, multi-head attention, the autoregressive layer, or singular value decomposition.

The experimental results indicate the following: 1) BiLSTNet-MHA demonstrates enhanced tracking of crown settlement sensor WY-03-01 in quarters exhibiting pronounced seasonal mixed periodicity, while the surrounding rock pressure sensor SL-S1-01 exhibits greater variability in the second quarter of 2020 due to its distinct trend compared to other quarters; 2) Examining the fourth-quarter data, BiLSTNet-MHA, based on the fourth-quarter 2020 metrics averaged across sensors of the same structural-response type, achieves the smallest errors among the compared models. In comparison to LSTNet, the mean absolute error and root mean squared error for surrounding rock pressure decrease by 43.15% and 48.79%, respectively, and the R2 score is also higher than that of LSTNet; 3) Removing the recurrent-skip layer increases MAE from 3.406 to 5.048, and the removal of the self-attention layer increases MAE to 4.471, indicating that the removal of recurrent skipping, multi-head attention, or the autoregressive layer degrades forecasting performance; 4) The removal of singular value decomposition slightly improves the average ablation accuracy but greatly increases training time.

BiLSTNet-MHA is capable of more comprehensively capturing long- and short-term periodic patterns as well as non-periodic variations in tunnel monitoring series, while reducing computational complexity and improving sensitivity to outliers. This makes it a valuable tool for real-world, practical tunnel health monitoring.