Vibration-based structural anomaly detection in real-time with piezoelectric patches on a tension rod assembly

This paper introduces a novel framework for real-time, local-level condition assessment of structural members using vibration-based anomaly detection. The approach employs a pair of piezoelectric patches: one acting as an actuator to excite the structural member and another as a sensor to capture the resulting vibration response. These signals are processed by four different machine learning models: two supervised and two unsupervised. Supervised models are a one-dimensional Convolutional Neural Network (1D CNN) and a hybrid Long Short-Term Memory model (1D CNN-LSTM). The unsupervised models are a one-dimensional Convolutional Autoencoder (1D CAE) and a hybrid 1D CAE-LSTM. The supervised algorithm basically classifies each signal segment as either “undamaged” or “damaged”. By aggregating the classification outputs across multiple segments, a structural health index is derived to quantify deviations from the baseline response of an undamaged member. In the unsupervised models, anomaly detection is based on reconstruction errors, which compute a similar health index by measuring deviation from the undamaged baseline response. The developed methods have been experimentally validated on a tension rod assembly, with damage simulated by reducing the applied tension. In this setup, the rod is threaded through a hollow steel section, and tensile force is adjusted via wing nuts at both ends. The fully tightened assembly represents the “undamaged” state, while the loosened conditions are considered as the “damaged” states. The system effectively identifies and quantifies damage severity in real-time, generating a visual graph for intuitive tracking of structural health changes. These applications demonstrate the potential of this method for practical use in monitoring real-world structures, such as suspension bridge cables and prestressed concrete elements. Experimental results confirm that the health indices derived from the proposed method closely align with the actual damage severity applied to the assembly, highlighting their accuracy and reliability.