Structural Health Monitoring (SHM) is a critical safety concern in the construction sector. Bridges, high-rise buildings and tunnel systems emit micro-seismic signals over time due to material fatigue and external loads. These signals appear long before visible damage, and when correctly interpreted, potential collapses can be prevented. The starting point of this research was the high false positive rate of conventional threshold-based trigger systems and their lack of real-time interpretation.
For the deep learning component, I converted raw seismic waveforms into time-frequency images (spectrograms) and trained a 1D CNN + Bi-LSTM hybrid architecture. Working with time series required a different preprocessing mindset than image classification: the choice of sliding window size, overlap ratio and FFT parameters directly affected model performance. Trained on both simulated and real field data, the model was able to distinguish four different micro-seismic pattern classes (crack formation, sliding, ground vibration, background noise).
The motivation for incorporating LLM integration into the research was this: when detection systems produce raw classification output, field engineers struggle to interpret it. A GPT-based language model was fine-tuned to receive seismic class labels, intensity metrics and location information and produce natural language explanations and action recommendations. This allowed the system to produce output such as 'Sliding-type micro-seismic activity detected in the north corridor. No immediate threat anticipated, but a visual inspection is recommended within 48 hours' rather than simply 'Class 2 pattern detected'.
The publication process was itself an educational experience. Publishing on ResearchGate familiarized me with academic writing standards, citation management and preprint culture before peer review. Reviewer comments brought to light some deficiencies in the system architecture; we needed to add extra validation steps particularly around how to reduce the hallucination risk of the LLM component. This iterative revision process showed that academic writing is a living process and contributed significantly to the maturation of the research.