تشخیص آسیب سازه‌ای با استفاده از شبکه‌ عصبی عمیق و تحلیل زمان–فرکانس سیگنال‌های ارتعاشی

In recent years, machine learning, particularly deep learning, has gained considerable attention as an effective approach for structural damage detection and improving structural safety. This study presents a deep learning–based framework for the diagnosis of structural damage. Vibration data were obtained from the LUMO steel lattice tower in Germany under both healthy and damaged conditions and employed as the primary dataset. The raw vibration signals were transformed into time–frequency representations (scalograms) using the continuous wavelet transform (CWT), and the resulting images were used as inputs to a deep convolutional neural network based on the ResNet-18 architecture.

The results demonstrate that the ResNet-18 model, owing to its residual connections and strong feature-learning capability, effectively extracted damage-related patterns and achieved high classification performance, with an accuracy of 97.5%, an F1-score of 94.7%, and a recall of 90%. Furthermore, the use of Grad-CAM enhanced the interpretability of the network by highlighting the time–frequency patterns within the scalograms that contributed most significantly to the model’s classification decisions. This analysis provides valuable insight into how the model distinguishes between healthy and damaged states, while remaining focused on the feature domain rather than the physical location of the damage. Overall, the proposed framework can accurately and reliably identify structural damage and provides a robust foundation for the development of real-time, vibration-based structural health monitoring (SHM) systems.