Vehicle Speed Invariant Drive-By Bridge Damage Detection

Abstract

Given the central role of bridges in transportation networks, continuous monitoring of these structures is crucial to detect damage, ensure long-term serviceability, and prevent catastrophic failures. Traditional inspection methods, however, are costly, laborintensive, and often subjective. Sensor-based approaches, such as strain gauges or accelerometers installed directly on bridges, require significant installation and maintenance efforts, limiting their scalability. An emerging alternative leverages vehicle–bridge interaction: by analyzing dynamic responses recorded by in-vehicle sensors, bridge conditions can be assessed without installing dedicated instrumentation on the structure. Vehicle accelerations reflect bridge-induced vibrations during crossings and indirectly encode dynamic properties of the bridge. However, variability in vehicle speed poses a significant challenge, as it affects the vibration signatures captured by accelerometers while remaining independent of bridge health. This confounding factor hinders the extraction of reliable, damage-sensitive features. To overcome this challenge, we propose a vehicle speed invariant drive-by bridge damage detection model that employs adversarial learning to extract features from vehicleinduced vibrations that are sensitive to structural damage while invariant to vehicle speed. The model integrates long short-term memory (LSTM) layers with a Gradient Reversal Layer (GRL). The LSTM layers capture temporal patterns in the frequency components, enabling the extraction of features that reflect subtle temporal variations across the full vibration spectrum. The GRL imposes speed invariance by adversarially optimizing the feature representation to maximize accuracy in damage classification while minimizing performance in speed prediction. We evaluate our method through a lab-scale experimental vehicle–bridge system with the vehicle running at varying speeds. Our model performs 1.4× better than baseline methods for bridge damage detection, achieving an accuracy of 91.01% and an F1-score of 93.29%.