Unsupervised Structural Damage Detection and Localization Using ۱D CNN Autoencoders on Vibration Data

Structural Health Monitoring (SHM) plays a vital role in ensuring the integrity and safety of critical infrastructure. However, supervised learning techniques often rely on labeled damage data, which can be costly, time-consuming, and impractical to obtain in real-world scenarios. To address this limitation, this study proposes an unsupervised deep learning approach for damage detection and localization in civil structures using vibration signals. The core idea is to train a ۱D Convolutional Neural Network (CNN)-based autoencoder solely on healthy structural data. The model learns to reconstruct normal vibration patterns, and damage is inferred from deviations in reconstruction accuracy when processing unseen signals. The proposed method processes raw vibration data collected from multiple sensors installed on a benchmark bridge structure. These signals are segmented into fixed-length windows and normalized before being fed into the autoencoder. During testing, the model computes reconstruction error for each signal segment. Anomalous segments are identified using a statistically derived threshold, typically based on the distribution of errors observed in healthy data. Damage localization is achieved by analyzing the spatial and temporal distribution of reconstruction errors across sensors and time, enabling the identification of both the presence and approximate location of structural anomalies. The experimental evaluation is conducted using the IASC-ASCE benchmark dataset, where the model is trained exclusively on undamaged cases and tested on various damage scenarios. The results demonstrate the model's ability to distinguish between healthy and damaged states with high reliability, without requiring any labeled damage data. Visualization tools such as reconstruction error histograms, time-series error plots, and sensor-wise anomaly heatmaps further support the model's interpretability and diagnostic capability. Overall, the approach shows promise for deployment in real-time SHM systems where early detection and localization of structural damage are critical.

کلیدواژه ها:

Damage Localization ، Unsupervised Learning ، Autoencoders ، vibration signal