Structural Health Monitoring Impact Classification Method Based on Bayesian Neural Network

Area of Science:

• Structural Health Monitoring (SHM)
• Composite Materials
• Machine Learning for Engineering

Background:

• Passive sensing for impact characterization often relies on deterministic methods, which struggle with real-world variability in impact conditions.
• Uncertainty in impact diagnosis arises from variations in location, angle, and energy, necessitating reliability-based approaches.

Purpose of the Study:

• To propose a novel, reliability-based impact characterization method using a Bayesian neural network (BNN) for multi-class classification and uncertainty quantification.
• To evaluate the robustness and reliability of the BNN against impact variability, including angled impacts.
• To compare the BNN's performance, uncertainty quantification, and computational efficiency against a multi-artificial neural network (multi-ANN).

Main Methods:

• Acquisition of impact data using a piezoelectric (PZT) sensor network on a composite plate.
• Feature extraction from sensor signals, including transferred energy, frequency at maximum amplitude, and time interval of the largest peak.
• Development and validation of a BNN model for classifying impact energy levels and quantifying diagnostic uncertainty, with comparative analysis against a multi-ANN.

Main Results:

• Both BNN and multi-ANN demonstrated high performance (94% and 98% reliable predictions, respectively) for classifying perpendicular impacts.
• Both models struggled with angled impacts not included in the training data, though uncertainty quantification provided additional diagnostic information.
• The BNN significantly outperformed the multi-ANN in terms of computational time and resource utilization.

Conclusions:

• The proposed BNN method provides a reliable approach for impact classification and uncertainty quantification in SHM, outperforming multi-ANN in computational efficiency.
• While effective for known impact conditions, further research is needed to improve model accuracy for novel scenarios like angled impacts.
• Uncertainty estimates from the BNN are valuable for interpreting diagnostic confidence and can guide future classification improvements.