Automatic Sensor Subset Selection for Vibration-Based Structural Damage Identification Using Multi-Task Deep Learning

This study proposes a sensor-gated multi-task deep-learning framework for automatically selecting compact sensor subsets for vibration-based structural damage identification. The framework jointly identifies damage locations and estimates stiffness-reduction ratios using natural-frequency changes and sensor-dependent modal ordinates. A physics-informed dataset containing 11551 structural states was generated using a frequency-calibrated reduced-order model of the Nam Xay Bridge, including one undamaged state, 1050 single-damage cases, and 10500 two-location damage cases. Seven candidate sensors, four vibration modes, 21 damage locations, and severity levels from 1% to 50% were considered. The learned sensor subsets were evaluated against all 127 non-empty combinations and conventional selection strategies. The five-sensor subset achieved the best overall performance, with a macro F1-score of 0.7620±0.0073 and a damage-severity MAE of 0.0705±0.0014, outperforming the seven-sensor reference while reducing the number of measurement channels by 28.6%. A more economical four-sensor configuration reduced the sensor count by 42.9%, with only a 1.1% decrease in macro F1-score relative to the full system. Under 10% input noise, its macro F1-score decreased by 2.2%, while the severity MAE increased by 5.0%. Ablation results confirmed that natural-frequency changes complemented the spatial information provided by modal ordinates. The results demonstrate the potential of joint sensor selection and damage identification, although validation using full finite-element and field measurements is required before bridge-specific deployment.