Better Structural Monitoring Using LSTM Neural Networks

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.May 14 2024

In a paper published in the journal Scientific Reports, researchers tackled sensor maintenance issues in structural monitoring systems by employing a long-short-term memory (LSTM) neural network to recover missing stress data during grid structure jacking construction.

Their method of analyzing data autocorrelation and spatial correlations proved effective. Comparative assessments favored the LSTM-based approach for its superior data restoration accuracy. Validation was done using Qingdao Jiao-dong International Airport's Hall F monitoring data during typhoon conditions.

Related Work

Previous research in structural health monitoring has explored diverse methods for repairing missing data, including Bayesian probability analysis, machine learning (ML) techniques, and correlation-based approaches. Deep learning (DL) methods, such as convolutional and recurrent neural networks (RNN), have emerged as a potent tool for data reconstruction.

Recent efforts have focused on integrating deep learning with Bayesian methods and developing hybrid models to address various data repair challenges effectively. These approaches have demonstrated promising results in reconstructing missing data and predicting structural responses across different monitoring scenarios.

Structural Data Repair

This paper introduces data repair methods tailored to various data loss scenarios encountered in structural monitoring systems, utilizing deep learning techniques and correlation analysis. Specifically, the proposed approaches address isolated data loss at single measurement points and data loss across multiple correlated points.

A self-correlation repair method is employed for isolated data loss, utilizing an LSTM model trained on existing monitoring data to predict and repair missing segments. When multiple correlated measurement points are affected, highly correlated data points are selected to train the LSTM model for repair. The process involves constructing datasets, training the LSTM neural network, and analyzing errors between predicted and missing values to evaluate model performance.

This study elaborates on the Pearson correlation coefficient method used for data correlation analysis and introduces the LSTM neural network as the primary tool for missing data imputation. The LSTM model consists of LSTM memory units and a fully connected layer, facilitating the processing of standardized input data sequences.

During training, the Adam optimization algorithm minimizes the loss function, ensuring optimal network structure. Model evaluation criteria, including root mean square error (RMSE) and mean absolute percent error (MAPE), are defined to assess the effectiveness of data repair, with smaller values indicating better performance.

Structural Response Repair

The team uses stress monitoring data from a steel structural truss roof system in grid jacking projects to explore the efficacy of structural response repair methods. The investigation focuses on a maintenance workshop characterized by a large-span steel structure consisting of orthogonal inclined trusses spanning 93 m in width and 65 m in length. 

A structural monitoring system comprising vibrating wire stress sensors and displacement measurement points was installed to capture internal force evolution during construction. Data selection and model training involved utilizing stress data from specific measurement points for missing data investigation, with a spatial correlation approach adopted for multiple measurement points. The LSTM model demonstrated robust performance in repairing missing data, with training loss convergence and accurate fitting results observed, indicating its suitability for structural health monitoring data repair.

Comparison of different missing data repair models revealed the superior performance of LSTM-based methods over traditional prediction models like support vector machines and multilayer perceptron (MLP) neural networks. Both LSTM-based approaches exhibited higher accuracy, with the spatial high correlation model showing slightly better repair accuracy than the self-correlation model. These findings underscore the effectiveness of LSTM neural networks in capturing temporal data features and highlight their suitability for structural health monitoring data repair in grid construction projects.

Terminal F Data Imputation

The focus lies on data imputation within the structural health monitoring system of Terminal F at Qingdao Jiao-dong International Airport, tackling missing data across stress, displacement, and vibration measurements critical for structural safety. Detailed descriptions of the terminal's layout, monitoring system composition, and data selection during a typhoon event underscore the necessity of addressing missing segments in crucial operational conditions.

With LSTM model training at a 20% loss rate, the effectiveness in recovering diverse monitoring data types is evident from convergence curves and recovery results, further validated by root mean square error (RMSE) and mean absolute percent error (MAPE)  metrics. These findings emphasize the robustness of LSTM-based data recovery, ensuring reliable structural safety assessment, particularly in critical scenarios like typhoon events, and demonstrating its practical applicability in complex infrastructure maintenance.

Conclusion

In summary, this paper introduced two LSTM-based methods for recovering missing stress data in large-span spatial structure monitoring systems. One method addressed isolated measurement points with self-correlation, while the other targeted multiple measurement points with high correlation.

Results demonstrated accurate stress reconstruction with MSE around 0.6 and MAPE below 15%, indicating superior performance to traditional prediction models like support vector machines and MLP neural networks. The method leveraging multiple measurement points with high correlation exhibited slightly higher accuracy than the self-correlation approach.