Sensor Placement and Measurement Fault Effects on Vibration-Based Structural Health Monitoring for Aircraft Wing Structures

Abstract

Sensor placement and measurement imperfections are examined jointly for vibration-based structural health monitoring of an aircraft-wing analog, using a simplified Euler-Bernoulli cantilever-beam model. Three placement strategies, Uniform, Antinode, and Effective Independence, are compared under a fixed five-sensor budget, with detection evaluated by a classical Mahalanobis-curvature detector and a Random Forest (RF) classifier trained on modal features. Both detectors are tested under clean conditions and four realistic sensor faults, additive noise, bias, dropout, and complete failure, using repeated random-split (K=10) and leave-one-damage-location-out (LODO) validation. Under clean conditions, the RF classifies all three placements’ 48-case test sets perfectly (accuracy = 1.000); under LODO, accuracy on withheld damage locations falls to 0.938 (SD = 0.125), showing strong but imperfect spatial generalization to unseen damage locations. The classical detector reaches 0.62-0.64 accuracy under clean conditions and collapses to chance (about 0.50) at the mildest tested additive-noise level; under bias and dropout, it retains a partial or placement-dependent signal at the mildest levels (0.602-0.633 at 1% bias and up to 0.548 at 10% dropout) before reaching chance-level performance at greater severities. The RF, trained with additive-noise augmentation but not bias, dropout, or failure corruption, degrades gradually under noise (to 0.54-0.57 at SNR = 10 dB) and bias (to 0.73-0.75 at 10% bias); under the mean-imputation and sensor-zeroing procedures used here, accuracy remains at or above 0.998 throughout the tested dropout and complete-failure conditions. Placement strategy has little effect on RF performance and does not consistently favor any classical-detector ordering under repeated sampling. Under the modeled conditions, the RF therefore sustains substantially higher absolute detection accuracy than the fault blind classical baseline, although the comparison is not fully equal because RF training includes additivenoise exposure whereas classical-detector calibration includes only healthy ambient-noise data.