TL;DR
This paper introduces Informer-SEnet, a novel deep learning model for predicting long-term structural responses in bridges. By combining the Informer's long-sequence forecasting capabilities with the Squeeze-and-Excitation (SE) attention mechanism, the model significantly improves prediction accuracy and robustness, crucial for bridge health monitoring and safety.
What Problem Does This Paper Address?
Bridges are complex structures whose responses to environmental and operational loads are stochastic, nonlinear, and highly coupled. Accurately predicting these responses over the long term is vital for effective bridge health monitoring (BHM), enabling timely anomaly detection, early damage warnings, and informed maintenance decisions. Traditional methods struggle with the high dimensionality, non-stationarity, and complex interdependencies present in multivariate time-series data from BHM systems. Existing deep learning models, while powerful, often face challenges with long-range dependency capture, feature redundancy, and computational efficiency when dealing with such data.
Methods
The core of this work is the Informer-SEnet model, which builds upon the Informer architecture. Key innovations include:
- Squeeze-and-Excitation (SE) Module: Integrated into the Informer framework, this module adaptively recalibrates the importance of different feature channels. It compresses spatial/temporal information (Squeeze) and then learns inter-channel dependencies to assign weights (Excitation), effectively highlighting crucial structural response features and suppressing redundant ones.
- Informer Architecture Enhancements:
- ProbSparse Self-Attention: Reduces computational complexity for long sequences by selectively focusing on the most important attention weights.
- Distillation Mechanism: Employs convolutional and pooling operations to compress time dimensions, extract multi-scale temporal features, and remove redundancy.
- Generative Decoder: Predicts the entire future sequence in a single forward pass, improving efficiency.
The paper also details data preprocessing steps, including handling anomalous data using ARIMA, Min-Max normalization for data scaling, and Variational Mode Decomposition (VMD) for separating temperature effects from strain data.
Datasets
The study utilized three datasets:
- Public Dataset: A standard open-source dataset commonly used in structural health monitoring research (Bridge-health-monitoring-master), containing strain, displacement, and temperature data.
- Measured Dataset 1: Real-world data from strain sensors at section 6 of an in-service bridge.
- Measured Dataset 2: Real-world data from strain sensors at section 5 of the same in-service bridge.
All datasets were preprocessed, and a sliding-window strategy was used to construct input-output pairs for training, validation, and testing. The prediction task involved using 24 hours of past data to predict the next 24 hours of temperature-induced strain.
Key Insights and Ideas
- Channel Attention is Key for Multivariate BHM: The SE module's ability to dynamically weigh feature channels proved crucial for effectively capturing the complex interplay of factors influencing structural responses. It's not just about temporal patterns but also about understanding which sensor readings are most informative at any given time.
- Encoder's Irreplaceable Role: Ablation studies highlighted the critical importance of the Informer's encoder, particularly its sparse self-attention and convolutional pooling mechanisms, for capturing long-range dependencies and local features essential for accurate long-sequence modeling.
- Robustness Through Stability: The Informer-SEnet model demonstrated significantly smaller fluctuations across cross-validation folds compared to baseline models. This indicates a more stable and reliable prediction capability, which is paramount for safety-critical applications like bridge monitoring.
- Beyond Numerical Gains: The reported improvements in MAE, RMSE, and R² translate directly into more precise characterization of structural behavior, leading to better anomaly detection, fewer false alarms, and more effective maintenance planning.
Key Results
The Informer-SEnet model achieved state-of-the-art performance across all tested datasets:
- Performance Improvements: In the best cases, the model achieved up to a 54.67% reduction in MAE, 52.39% reduction in RMSE, and a 7.73% improvement in R² compared to baseline models.
- Comparative Experiments: Consistently outperformed traditional RNNs (LSTM, BiLSTM) and Transformer variants (Transformer, Transformer-Encoder, Informer, Informer-Encoder) on both public and real-world bridge datasets.
- Ablation Studies: Confirmed the significant contributions of the SE module, sparse attention, and distillation mechanisms to the overall performance. Removing any of these components led to a noticeable degradation in prediction accuracy and stability.
- Cross-Validation: Demonstrated superior robustness and generalization ability, with smaller standard deviations in performance metrics across different data partitions.
Limitations and Future Work
- Limitations: While the paper highlights the model's strengths, it implicitly acknowledges that capturing extremely complex, multi-source coupled monitoring data remains a challenge. The "multi-modal" error characteristics of baseline models suggest they struggle with sudden changes and non-stationary behavior, a problem Informer-SEnet mitigates but may not entirely eliminate.
- Future Work: The paper suggests that the Informer-SEnet model provides an "engineering-ready tool" for intelligent bridge operation and maintenance. Future work could focus on deploying this model in real-time BHM systems, exploring its application to other types of structural health monitoring (e.g., tunnels, buildings), and further investigating its capabilities in detecting specific types of structural damage. The authors also mention the potential for further optimization of the SE module and attention mechanisms.
Summary updated