Leveraging AI for Structural Health Monitoring: Ultrasonic Guided Waves in Predicting Delamination Damage in Aircraft Composites

11 Leveraging AI for Structural Health Monitoring: Ultrasonic Guided Waves in Predicting Delamination Damage in Aircraft Composites Panagiotis Kolozis1, José Manuel Royo2, Javier Hernandez-Olivan2, Vanessa Thalassinou-Lislevand1, Andrea Calvo Echenique2, Elias Koumoulos1 1. Innovation in Research and Engineering Solutions, Schaerbeek, Belgium 2. Technological Institute of Aragon, Zaragoza, Spain 8TH ICEAF Conference 22-24 June 2025, Kalamata Greece This research was funded by the European Union’s Horizon 2020 Research and Innovation Programme GENEX under grant number 101056822 2 •Composite materials are used in aviation for their high strength-to-weight ratio. •However, they are vulnerable to hidden defects, such as delamination. •Ensuring structural integrity is critical for safety, efficiency, and cost reduction. The Role of Structural Health Monitoring (SHM) •Technologies Used: Piezoelectric transducers, Acoustic emission detectors •Benefits: Early failure detection, lower maintenance costs, enhance safety and sustainability. Motivation & Context 3 Develop an AI-powered SHM method to: •Detect and localize delamination in composite laminates. •Predict damage growth using sensor-based data. •Bridge the gap between simulated and real-world measurements. Method: •Ultrasonic Guided Wave (UGW) simulations •Experimental testing with piezoelectric sensors •Deep learning for signal analysis and noise mitigation Research Objectives 4 The Design of Experiment for the FE simulations provided by is defined as follows: •Variables of the DoE: location of damage in x, y axis, damage size in x, y axis •Four batches of simulations of different damage sizes: 1. The damage size in the x, y axis of each batch was (0.12m, 0.08m), (0.09m,0.06m), (0.06m,0.04m) and (0.03m, 0.02m) respectively. 2. Each batch had simulations of 0.01m spacing in damage locations, 24 (16) positions for the x-axis (y-axis). •All the defects were simulated at the third quadrant of the plate •Automatically generate the data from defects in all quadrants by applying relevant symmetries Data generation 0.32m 0.48m 5 Feature engineering steps 1. Signal smoothing – 1D interpolation (time domain) 2. Pristine subtraction 3. Estimate signal energy (sum of squares) in 5-time windows (5 features per sensor) 4. Fast Fourier transformation (Normalized to unity to emphasize on shape differences) 5. Estimate “energy” (sum of squares) of normalized FFTs in 5 frequency windows (+5 features per sensor) Feature engineering 6 The data were fed into a Deep learning model with sequential linear layers •The loss function was defined as the Mean Square Error between the true and predicted values for the damage location and sizes •The performance of the model was validated on a “test” subset of data that was not used during training Model training & validation Model training & validation features targets 90 features 4 targets Linear Model architecture: dmg_loc Distance prediction performance Dmg size prediction performance Overlap percentage Loss function vs epoch 7 Test the performance on datasets simulated at a different temperatures that could serve as a more realistic validation set The simulations had a constant damage size of (0.12, 0.08) on 5 different locations across the laminate and temperatures 10, 15, 18, 20, 22, 25, 30 Celsius degrees Datasets with simulated noise due to temperature variation In the plots, the same sensor signal can be seen for the initial simulation with red and other colours for different temperatures 8 •This model predicts with good accuracy the damage location and size •This indicates that the engineered features have some robustness to this kind of noise Performance validation on noisy data Distance prediction performance Dmg size prediction performance Overlap percentage 9 Problem Statement: Even with sensors placed in the same nominal location, signal footprints can vary from plate to plate due to: •Material inhomogeneities •Sensor placement precision •Variability in sensor properties •Attachment quality between sensor and plate Experimental Approach: •Four identical composite plates tested •Same sensor network, with defects at different initial positions •Focus: Delamination progression Our Hypothesis: •Perfect time-domain matching between simulation & experiment is not feasible •But: Trends in signal features due to damage growth are consistent across plates •Goal: Characterize these damage-correlated variations, not exact signals Modeling Experimental Uncertainty in SHM using Machine Learning