Abstract
<jats:p>Detecting incipient structural damage in civil infrastructure requires isolating transient acoustic emissions that are heavily obscured by macroscopic environmental noise. This paper introduces SPECTRA-CQ (Spectrogram Pattern Evaluation via Classical and Quantum Architectures), a hybrid anomaly detection framework evaluated on continuous vibration telemetry from the operational openLAB reference bridge in Bautzen, Germany. Raw 1-dimensional accelerometer signals are mapped into 2-dimensional 64 × 1000 energy-density spectrograms via Continuous Wavelet Transform (CWT) to preserve critical time-frequency localization. A synthetic acoustic emission (AE) proxy is then injected into a target node to establish strict physical detection limits. This AE is bounded at an amplitude of 0.35 to mathematically simulate early-stage micro-cracking rather than obvious macroscopic failure. Initial unconstrained baseline ablations establish that both architectures are mathematically sound. Both Classical Convolutional Autoencoders (CAE) and Hybrid Quantum Autoencoders (HQAE) easily capture the structural variance at a latent space of 8 qubits on a single sensor and yield highly accurate reconstructions with a mean squared error of approximately 0.008. A fundamental bottleneck emerges when enforcing Noisy Intermediate-Scale Quantum (NISQ) capacity constraints. Aggressively compressing the latent space to just 2 qubits forces isolated single-sensor models to fail. Neither the classical nor the quantum architecture can effectively separate low-amplitude anomalies from standard reconstruction noise without a broader spatial context. This limitation, called “Single-Sensor Spatial Blindness” prevents reliable incipient damage detection. This barrier is overcome by expanding the architecture into a 6-sensor Siamese convolutional graph. It fuses spatial data from all six sensors while strictly maintaining the two-qubit bottleneck for each specific sensor. This setup allows the model to anchor itself against the geometric state of adjacent healthy nodes. This multi-sensor fusion successfully recovers localized detection for the classical CAE architecture and scales to an AUC-ROC of 0.888 for incipient damage evaluated at higher spectral mask widths. The quantum equivalent, however, hits a hard scaling wall. Ablation studies on circuit topology investigate the exact cause of this collapse. Standard parameterized quantum circuits, ranging from completely uncoupled rotations to fully entangled arrays, are evaluated against two custom entanglement topologies. These custom designs include a rigidly structured global mapping and a Physics-Informed Neural Network (PINN) configuration designed to mathematically embed continuous wave attenuation. Despite these targeted configurations, forcing global entanglement across 12 qubits still reliably triggers gradient vanishing in the current NISQ-era hardware. Heavily entangled spatial quantum networks succumb to barren plateaus and degrade incipient detection back to baseline levels. Ultimately, the data highlights a clear divergence. Classical architectures readily exploit spatial coherence to bypass extreme latent compression. Forced quantum entanglement currently acts as a strict mathematical barrier for continuous multi-sensor structural monitoring.</jats:p>