Numerical Analysis of Experimental Uncertainties in Ultrasonic Guided Waves Propagation for Damage Monitoring in Composite Structures

Numerical Analysis of Experimental Uncertainties in Ultrasonic Guided Waves Propagation for Damage Monitoring in Composite Structures Javier Hernandez-Olivan, Panagiotis Kolozis, José Manuel Royo, Andrea Calvo-Echenique, Susana Calvo, Elias P. Koumoulos Predictive Maintenance in the Aerospace Industry 2 This work was supported by the European Commission [grant agreement No 101056822]. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union. Experimental coupons Numerical models ULTRASONIC GUIDED WAVES (UGW) Validation at coupon level Solution proposed A ML-based SHM system to predict delamination propagation in CFRP plates using UGW Composite materials have been introduced in the aerospace industry due to their high stiffness to weight ratio - Failure modes not visible to the naked eye - Inefficient and costly traditional maintenance - Structural Health Monitoring methodologies - Damage prediction with numerical validation - Digital Twins updated in real-time Predictive Maintenance Condition-based Maintenance The need to compensate the numerical-experimental uncertainty 3 01 Introduction The Validation Methodology: From Inspection to Prediction Monitoring Campaign 02 Materials and methods Experimental Set-up Numerical Model Feature Extraction Method Numerical-Experimental Validation Table of Contents 03 Machine Learning (ML) Strategies Why Machine Learning (ML) ? ML Flow Chart ML Model with Different Specimens 04 Results Random Forest Evaluation in P1 from 75% to 5% train data Strategy 1: ML algorithm first IT train rest ITs test all plates Strategy 2: ML algorithm one plate train rest plates test 05 Conclusions Conclusions and future work Introduction 01 5 Prognosis Physics-based ML algorithms and ROMs Scalable multiphysics simulations Replicate experimental features Stepwise Circular Flexion Loading Test Thermography UGW Monitoring Experimental Campaign Numerical Model Validation ML Model Predictive Monitoring Fracture Test Monitoring Test Abaqus Simulations Signal Processing Feature Extraction Feature Prediction Condition Monitoring Predictive Monitoring Sensor network optimization The Validation Methodology: From Inspection to Prediction ◼Objective: Study how to compensate the numerical-experimental uncertainties of the propagation of ultrasonic guided waves (UGW), generated by directional sensors (MFC), in an orthotropic material damaged by delamination. ◼Material: Four Carbon Fiber Reinforced Polymer (CFRP) laminates 500x250x3.93 mm (LxWxT) were manufactured at AIMEN by Automatic Fiber Placement (AFP) and instrumented and tested at ITA under stepwise circular flexion loading test. ◼Sensors: Eight Macro Fibre Composite (MFC) sensors (M-0714-P2) commercialized by Smart Materials, directly attached to the top surface of the plate with cyanoacrylate. ◼UGW acquisition system: Phased Array Monitoring for Enhanced Life Assessment (SHMUS) as UGW generator. (Aranguren et al., 2018) ◼Other inspection techniques: Thermography and Penetrating Liquids, to locate and quantify damage. 6 Monitoring Campaign Load Precrack Support Teflon Stacking Sequence 21 plies: (45/-45/0/-45/45/90/-45/45/45/-45/90/ T/-45/45/45/-45/90/45/-45/0/-45/45) E1[Pa] E2[Pa] E3[Pa] n12 n13 n23 G12 [Pa] G13 [Pa] G23 [Pa] ρ [kg/m3] 4.25E+10 3.79E+10 1.51E+10 0.5721 0.2224 0.2425 3.04E+09 5.20E+09 5.20E+09 1.586E+03 Materials and methods 02 8 UTM Cross-head CFRP plateActuator Support Actuator Support Stepwise Circular Flexion Loading Test Set-up Stepwise Circular Flexion Loading Test Set-up Ultrasonic Guided Wave (UGW) Set-up Thermography Set-up Thermal imaging camera CFRP plate Heat source (light bulb) Thermal image Delamination Input signal: Hanning window 3.5 cyclesinusoidal signal, 20-V amp at 150 kHz. MFC M-0714-P2 SHMUS Jack audio connection cables Experimental Campaign 9 Numerical Model ◼Numerical purpose: Validate models for generating synthetic datasets for data augmentation. ◼Before modelling: 1. Locate and quantify the damage. 2. Check geometry, material properties, excitation signal, sensor placement and boundary conditions. 3. Obtain the dispersion curves, to calculate the element length, and identify waves modes. ◼Modelling procedure: 1. Mesh: Plate C3D8, 1-mm length. Sensor C3D8E, 1-mm length. 2. The transducers are attached to the plate by shared nodes between their respective meshes. 3. Damage: Two plates of 1/2 thickness are created. Abaqus "Tie constrain" merges coincident nodes. Delamination results in a free surface. 4. Co-simulation step: Abaqus Standard (implicit strategy) for piezoelectric behaviour and Abaqus Explicit (explicit strategy) for elastic behaviour. Delamination (free Surface) Tie Constrain (same displacements) Results 04 Random Forest Evaluation in P1 from 75% to 5% train data 17 ◼Data: Total features = 10 features/signal * 8 signals/RoundRobin * 8 RoundRobin/iteration * 4 iterations = 2560 features 1. Signal features: signal energies and FFT energies in each signal window. 2. Damage characteristics: damage coordinates (X, Y, Z) and damage size. 3. Train data: from 75% to 5% of total data (1920 –640 samples). 4. Test data: a fixed 20% of total data (512 samples). 5. Repeat the process 10 times (10 Runs) 25% train data 640 train samples (features) 64 signals + BL (64 signals) 8 Round Robin + BL (8 RR) 1 iteration + BL iteration 75% train data 1920 train samples (features) 192 signals + BL (64 signals) 24 Round Robin + BL (8 RR) 3 iterations + BL iteration Strategy 1: ML Algorithm First IT train Rest ITs test ALL PLATES 18 Model N_Train N_Test MAE R2 Percent_In_Band Time_s Best_Params Random Forest 2560 6400 0.03982 0.93199 87.71875 6.82 max_depth : 10, min_samples_split: 2, n_estimators: 100 Gradient Boost 2560 6400 0.04588 0.92226 85.5 2.82 learning_rate : 0.05, max_depth: 3, n_estimators: 100 XGBoost 2560 6400 0.04322 0.93181 86.9375 0.51 learning_rate : 0.1, max_depth: 3, n_estimators: 100, subsample: 0.8 do o v d d oo v d d oo oo v d do o v d Strategy 2: ML Algorithm One Plate train Rest Plates test 19 Model R2 Average (Stand. Dev.) MAE Average (Stand. Dev.) % Band Average (Stand. Dev.) N Train (P1 Train) N Test (Rest P Test) Total Time Average (s) Random Forest 0.8041 ( ±0.0390) 0.0735 ( ±0.0064) 70.9590 ( ±3.2555) 2560 6400 11.00 Gradient Boost 0.8226 ( ±0.0340) 0.0707 ( ±0.0072) 72.4766 ( ±3.5069) 2560 6400 6.10 XGBoost 0.7937 ( ±0.0461) 0.0784 ( ±0.0117) 69.7148 ( ±5.2023) 2560 6400 1.42 d oo v d d oo oo v d Conclusions 05 ◼Delamination monitorization has been done in CFRP laminates experimentally and numerically. ◼A feature extraction method has been utilized to extract signal features in both time and frequency domain. ◼Experimental uncertainties can complicate a numerical-experimental validation. Explore signal features instead of signal forms. ◼It is possible to predict experimental features from numerical features, without having a numerical-experimental validation. ◼Predict future damage states from only the first state: 1. Best results (MAE, R2, % data in band):Random Forest > Extreme Gradient Boost > Gradient Boost 2. Minor Time training: Extreme Gradient Boost < Gradient Boost < Random Forest ◼Predict damage states of other plates from a single plate: 1. Best results (MAE, R2, % data in band): Gradient Boost > Random Forest > Extreme Gradient Boost 2. Minor Time training: Extreme Gradient Boost < Gradient Boost < Random Forest Future Work: Damage detection, localization and quantification algorithm with the feature extraction method integrating the ML model. Conclusions 21 22 Thank you! Presenter: [email protected]