Sensor Data Reconstruction for Dynamic Responses of Structures Using External Feedback of Recurrent Neural Network

Area of Science:

• Structural Engineering
• Data Science
• Artificial Intelligence

Background:

• Sensor faults in structural health monitoring systems can compromise data integrity and hinder accurate structural condition assessment.
• Data reconstruction techniques are crucial for restoring datasets with missing sensor information.

Purpose of the Study:

• To propose a novel recurrent neural network (RNN) model for enhancing the accuracy and effectiveness of sensor data reconstruction.
• To improve the measurement of dynamic structural responses despite sensor faults.

Main Methods:

• A recurrent neural network (RNN) model incorporating external feedback was developed.
• The model leverages spatial correlation by feeding previously reconstructed data of defective sensor channels back into the input.
• Various RNN architectures, including simple RNN, LSTM, and GRU, were trained and validated on acceleration datasets from scaled building frames.

Main Results:

• The proposed RNN model demonstrated robust and precise sensor data reconstruction.
• The method's performance was found to be largely independent of RNN hyperparameters due to its reliance on spatial correlation.
• Validation on laboratory-scaled shear building frames confirmed the model's effectiveness.

Conclusions:

• The proposed RNN-based sensor data reconstruction method effectively addresses challenges posed by sensor faults in structural health monitoring.
• Utilizing spatial correlation enhances the robustness and accuracy of dynamic response measurements.
• This approach offers a reliable solution for maintaining the integrity of structural condition assessment.