Incipient Anomaly Detection in Bridge Health Monitoring using Capacity-Constrained Classical and Hybrid-Quantum Autoencoders

Abstract

This paper introduces a vibration-based machine learning framework for anomaly detection in bridges. Structural signals need to be isolated from stochastic noise for incipient damage detection in civil infrastructure. This study uses Continuous Wavelet Transform to map raw accelerometer telemetry from the openLAB reference bridge into 2D time-frequency domain to allow efficient damage localisation. The 2D matrices are passed through a convolutional frontend which employs pooling layers to filter environmental noise. An acoustic emission proxy is injected into a targeted sensor to simulate early-stage micro-cracking. Various classical and hybrid-quantum autoencoder topologies, including a physics-informed neural network, are designed and compared. All models share the same convolutional frontend and decoder; the hybrid quantum models replace the classical bottleneck layer with a variational quantum classifier. Note that all experiments were simulated on classical hardware using Pennylane.

Each sensor channel is constrained to a two-variable bottleneck to mirror current noisy intermediate-scale quantum hardware limitations. An isolated single-sensor architecture fails to differentiate damage transients from baseline noise and routine reconstruction errors. This limitation is overcome by expanding the architecture into a 6-sensor weight-tied Siamese convolutional network that maintains the strict bottleneck constraint while capturing geometric context of the bridge deck. Evaluation shows spatial fusion successfully recovers anomaly detection capacity for classical convolutional networks and unentangled quantum baselines. However, forcing inter-sensor quantum entanglement currently triggers gradient vanishing.