PSI - Issue 77

Douaa Benhaddouche et al. / Procedia Structural Integrity 77 (2026) 152–160 Douaa BENHADDOUCHE/ Structural Integrity Procedia 00 (2026) 000 – 000

157

6

Fig.4. Sensor Graph

Fig.3. DTW Distances over sensors

The heatmap of Fig. 3 shows the DTW distances over sensors at the afternoon. To ensure sufficient connectivity and preserve spatial relationships, the graph was further constrained so that each sensor had at least two edges, corresponding to its two most similar sensors. Consequently, a fixed threshold of 105 was applied to define similarity, such that an edge was created between two sensors if their DTW distance was below this value. The figure 3.b illustrates the sensor graph in the afternoon. The GCN-LSTM model was trained using CPU in 7 core machine, the architecture was built using TensorFlow. The model was trained on batch of 128 samples, the windows sequence for prediction was set to 12, learning rate to 0.0005 and LSTM units 64. These model hyperparameters were determined through experimental tuning. Before feeding into the model all data were normalized between -1 and 1. Morning Afternoon Evening (a) (c) Fig. 5. Correlations between the forecasted values and the real values for Morning, Afternoon and evening of sensors: (a) S1, (b) S7 and (c) S13. The figure 4 presents the correlation between the predicted and actual values of acceleration data of sensors S1, S7 and S13 from validation set (03/02/2008). The coefficient of determination (R²) was used to evaluate the model predictions. Across the three periods (morning, afternoon, and evening), the R² values for all sensors range between 0.95 and 0.98, indicating a high prediction accuracy. Moreover, most points lie close to the red reference line (real signal), with only weak dispersion, which reflects the small values of MSE. Notably, the evening period shows the best performance, which may be attributed to the fact that nighttime accelerations are generally more stable and easier (b)