# Structural Damage Identification of Composite Rotors Based on Fully Connected Neural Networks and Convolutional Neural Networks

^{1}

^{2}

^{3}

^{*}

## Abstract

**:**

## 1. Introduction

#### 1.1. State-of-the-Art

#### 1.2. Aim and Outline of the Paper

## 2. Materials and Methods

#### 2.1. Investigated Composite Rotors

#### 2.2. Data Set Creation and Data Augmentation

#### 2.3. Labeling of the Damage States

#### 2.4. Dimensionality Reduction Methods

#### 2.5. Model Training, Validation and Testing

#### 2.6. Overview of the Generated Machine Learning Models

- Algorithm: FC, 1D-CNN,
- Data set: original (OD) or augmented (AD),
- Dimensionality reduction method: pure data, each 30th frequency, PCA, PCA+LDA, SVD+LDA and KPCA,
- Classifier: radial position (RAD), angular position (ANG), load magnitude (MAGN) and damage accumulation (DAM).

## 3. Results

#### 3.1. Classification with Fully Connected Neural Network and Convolutional Neural Network

#### 3.2. Influence of Reduced Dimensionality

#### 3.3. Influence of Synthetically Augmented Data Set

## 4. Discussion and Conclusions

## Author Contributions

## Funding

## Institutional Review Board Statement

## Informed Consent Statement

## Data Availability Statement

## Acknowledgments

## Conflicts of Interest

## Abbreviations

ANG | Angular Position |

ANN | Artificial Neural Network |

CNN | Convolutional Neural Network |

DAM | Damage Accumulation |

FC | Fully Connected Neural Network |

FE | Finite Element |

FRF | Frequency Response Function |

HMM | Hidden Markov Model |

KPCA | Kernel Principal Component Analysis |

LDA | Linear Discriminanat Analysis |

MAGN | Load Magnitude |

ML | Machine Learning |

OD | Original Data |

PCA | Principal Component Analysis |

PSD | Power Spectral Density |

RAD | Radial Position |

ReLU | Rectified Linear Unit |

SHM | Structural Health Monitoring |

SVD | Singular Value Decomposition |

## Appendix A

**Table A1.**All trained models and combinations of network, data type and dimensionality. Test accuracy, sensitivity, specificity and validation accuracy with cross-validation error is displayed.

Network Type | Data Type | Dimensionality Reduction | Classifier | Name | Acc. [%] | Sens. [%] | Spec. [%] |
---|---|---|---|---|---|---|---|

FC | Original logarithmized data | Pure data | RAD | FC-OD-Pure-RAD | $99.2\pm 0.5$ | $99.4\pm 0.6$ | $98.9\pm 0.6$ |

FC | Original logarithmized data | Pure data | ANG | FC-OD-Pure-ANG | $32.7\pm 4.9$ | $32.9\pm 7.5$ | $34.6\pm 6.0$ |

FC | Original logarithmized data | Pure data | MAGN | FC-OD-Pure-MAGN | $94.0\pm 4.3$ | $99.0\pm 1.5$ | $91.1\pm 7.0$ |

FC | Original logarithmized data | Pure data | DAM | FC-OD-Pure-DAM | $87.1\pm 1.4$ | $96.2\pm 2.6$ | $83.4\pm 2.5$ |

FC | Original logarithmized data | Every 30th frequency | RAD | FC-OD-30thFreq-RAD | $97.6\pm 1.4$ | $99.2\pm 1.1$ | $96.1\pm 2.5$ |

FC | Original logarithmized data | Every 30th frequency | ANG | FC-OD-30thFreq-ANG | − | − | − |

FC | Original logarithmized data | Every 30th frequency | MAGN | FC-OD-30thFreq-MAGN | $97.1\pm 1.8$ | $98.4\pm 2.2$ | $96.0\pm 2.1$ |

FC | Original logarithmized data | Every 30th frequency | DAM | FC-OD-30thFreq-DAM | $87.6\pm 1.6$ | $92.3\pm 6.3$ | $85.9\pm 2.8$ |

FC | Original logarithmized data | PCA | RAD | FC-OD-PCA-RDA | $99.0\pm 0.7$ | $100.0\pm 0$ | $98.1\pm 1.3$ |

FC | Original logarithmized data | PCA | ANG | FC-OD-PCA-ANG | − | − | − |

FC | Original logarithmized data | PCA | MAGN | FC-OD-PCA-MAGN | $95.8\pm 1.6$ | $95.9\pm 2.7$ | $95.8\pm 2.7$ |

FC | Original logarithmized data | PCA | DAM | FC-OD-PCA-DAM | $87.6\pm 2.5$ | $94.2\pm 3.1$ | $84.5\pm 3.9$ |

FC | Original logarithmized data | PCA + LDA | RAD | FC-OD-PCA+LDA-RAD | $98.6\pm 0.6$ | $99.4\pm 0.7$ | $97.7\pm 1.3$ |

FC | Original logarithmized data | PCA + LDA | ANG | FC-OD-PCA+LDA-ANG | − | − | − |

FC | Original logarithmized data | PCA + LDA | MAGN | FC-OD-PCA+LDA-MAGN | $94.6\pm 2.0$ | $97.3\pm 2.0$ | $92.2\pm 2.1$ |

FC | Original logarithmized data | PCA + LDA | DAM | FC-OD-PCA+LDA-DAM | $84.2\pm 1.4$ | $93.5\pm 3.8$ | $80.3\pm 2.7$ |

FC | Original logarithmized data | SVD + LDA | RAD | FC-OD-SVD+LDA-RAD | $92.2\pm 2.5$ | $91.3\pm 2.7$ | $93.0\pm 3.8$ |

FC | Original logarithmized data | SVD + LDA | ANG | FC-OD-SVD+LDA-ANG | − | − | − |

FC | Original logarithmized data | SVD + LDA | MAGN | FC-OD-SVD+LDA-MAGN | $77.1\pm 4.5$ | $79.9\pm 3.4$ | $74.9\pm 6.9$ |

FC | Original logarithmized data | SVD + LDA | DAM | FC-OD-SVD+LDA-DAM | $73.6\pm 2.7$ | $89.3\pm 5.2$ | $67.6\pm 2.9$ |

FC | Original logarithmized data | KPCA | RAD | FC-OD-KPCA-RAD | $99.3\pm 6.2$ | $99.5\pm 1.0$ | $99.2\pm 1.0$ |

FC | Original logarithmized data | KPCA | ANG | FC-OD-KPCA-ANG | − | − | − |

FC | Original logarithmized data | KPCA | MAGN | FC-OD-KPCA-MAGN | $96.4\pm 1.5$ | $96.8\pm 1.5$ | $95.9\pm 2.5$ |

FC | Original logarithmized data | KPCA | DAM | FC-OD-KPCA-DAM | $89.2\pm 3.4$ | $94.3\pm 2.0$ | $86.9\pm 4.5$ |

FC | Augmented logarithmized data | Pure data | RAD | FC-AD-Pure-RAD | $99.5\pm 0.4$ | $99.7\pm 0.4$ | $99.3\pm 0.5$ |

FC | Augmented logarithmized data | Pure data | ANG | FC-AD-Pure-ANG | $73.9\pm 2.3$ | $81.4\pm 2.5$ | $71.1\pm 3.0$ |

FC | Augmented logarithmized data | Pure data | MAGN | FC-AD-Pure-MAGN | $97.3\pm 0.5$ | $96.7\pm 1.2$ | $98.0\pm 0.5$ |

FC | Augmented logarithmized data | Pure data | DAM | FC-AD-Pure-DAM | $96.8\pm 0.5$ | $98.5\pm 0.3$ | $96.1\pm 0.7$ |

1D-CNN | Original logarithmized data | Pure data | RAD | CNN-OD-Pure-RAD | $99.3\pm 0.6$ | $99.1\pm 0.9$ | $99.4\pm 0.6$ |

1D-CNN | Original logarithmized data | Pure data | ANG | CNN-OD-Pure-ANG | $32.6\pm 1.4$ | $32.4\pm 1.4$ | $32.9\pm 1.6$ |

1D-CNN | Original logarithmized data | Pure data | MAGN | CNN-OD-Pure-MAGN | $95.0\pm 2.0$ | $95.5\pm 4.2$ | $94.4\pm 3.9$ |

1D-CNN | Original logarithmized data | Pure data | DAM | CNN-OD-Pure-DAM | $87.2\pm 3.4$ | $97.0\pm 2.9$ | $83.3\pm 5.3$ |

1D-CNN | Augmented logarithmized data | Pure data | RAD | CNN-AD-Pure-RAD | $99.5\pm 0.2$ | $99.5\pm 0.3$ | $99.1\pm 0.2$ |

1D-CNN | Augmented logarithmized data | Pure data | ANG | CNN-AD-Pure-ANG | $70.7\pm 4.5$ | $68.8\pm 4.6$ | $71.7\pm 4.6$ |

1D-CNN | Augmented logarithmized data | Pure data | MAGN | CNN-AD-Pure-MAGN | $95.3\pm 0.6$ | $95.7\pm 1.2$ | $94.9\pm 1.6$ |

1D-CNN | Augmented logarithmized data | Pure data | DAM | CNN-AD-Pure-DAM | $94.9\pm 0.8$ | $97.3\pm 0.3$ | $93.8\pm 1.2$ |

## References

- Rafiei, M.; Adeli, H. A novel machine learning-based algorithm to detect damage in high-rise building structures. Struct. Des. Tall Spec. Build.
**2017**, 26, e1400. [Google Scholar] [CrossRef] - Agdas, D.; Rice, J.; Martinez, J.; Lasa, I. Comparison of visual inspection and structural-health monitoring as bridge condition assessment methods. J. Perform. Constr. Facil.
**2016**, 30, 04015049. [Google Scholar] [CrossRef][Green Version] - Kang, F.; Liu, J.; Li, J.; Li, S. Concrete dam deformation prediction model for health monitoring based on extreme learning machine. Struct. Control Health Monit.
**2017**, 24, e1997. [Google Scholar] [CrossRef] - Bossi, G.; Schenato, L.; Marcato, G. Structural health monitoring of a road tunnel intersecting a large and active landslide. Appl. Sci.
**2017**, 7, 1271. [Google Scholar] - Ciang, C.; Lee, J.; Bang, H. Structural health monitoring for a wind turbine system: A review of damage detection methods. Meas. Sci. Technol.
**2008**, 19, 122001. [Google Scholar] [CrossRef][Green Version] - Abdallah, I.; Dertimanis, V.; Mylonas, H.; Tatsis, K.; Chatzi, E.; Dervilis, N.; Worden, K.; Maguire, E. Fault diagnosis of wind turbine structures using decision tree learning algorithms with big data. In Proceedings of the Safety and Reliability-Safe Societies in a Changing World, Penang, Malaysia, 10–12 April 2018; Volume 2016, pp. 3053–3061. [Google Scholar]
- Dervilis, N.; Worden, K.; Cross, E. On robust regression analysis as a means of exploring environmental and operational conditions for SHM data. J. Sound Vib.
**2015**, 347, 279–296. [Google Scholar] [CrossRef][Green Version] - Vitola, J.; Pozo, F.; Tibaduiza, D.; Anaya, M. A sensor data fusion system based on k-nearest neighbor pattern classification for structural health monitoring applications. Sensors
**2017**, 17, 417. [Google Scholar] [CrossRef] [PubMed] - Santos, A.; Figueiredo, E.; Silva, M.; Sales, C.; Costa, J. Machine learning algorithms for damage detection: Kernel-based approaches. J. Sound Vib.
**2016**, 363, 584–599. [Google Scholar] [CrossRef] - Eleftheroglou, N.; Zarouchas, D.; Loutas, T.; Alderliesten, R.C.; Benedictus, R. Online remaining fatigue life prognosis for composite materials based on strain data and stochastic modeling. Key Eng. Mater. Trans. Tech. Publ.
**2016**, 713, 34–37. [Google Scholar] [CrossRef][Green Version] - Filippatos, A.; Langkamp, A.; Kostka, P.; Gude, M. A Sequence-Based Damage Identification Method for Composite Rotors by Applying the Kullback–Leibler Divergence, a Two-Sample Kolmogorov–Smirnov Test and a Statistical Hidden Markov Model. Entropy
**2019**, 21, 690. [Google Scholar] [CrossRef][Green Version] - Chetwynd, D.; Mustapha, F.; Worden, K.; Rongong, J.; Pierce, S.; Dulieu-Barton, J. Damage localisation in a stiffened composite panel. Strain
**2008**, 44, 298–307. [Google Scholar] [CrossRef] - Islam, A.; Craig, K. Damage detection in composite structures using piezoelectric materials (and neural net). Smart Mater. Struct.
**1994**, 3, 318. [Google Scholar] [CrossRef] - Watkins, S.; Akhavan, F.; Dua, R.; Chandrashekhara, K.; Wunsch, D. Impact-induced damage characterization of composite plates using neural networks. Smart Mater. Struct.
**2007**, 16, 515. [Google Scholar] [CrossRef][Green Version] - Dua, R.; Watkins, S.E.; Wunsch, D.C.; Chandrashekhara, K.; Akhavan, F. Detection and classification of impact-induced damage in composite plates using neural networks. In Proceedings of the IJCNN’01, International Joint Conference on Neural Networks, Proceedings (Cat. No. 01CH37222), Washington, DC, USA, 15–19 July 2001. [Google Scholar]
- Sammons, D.; Winfree, W.P.; Burke, E.; Ji, S. Segmenting delaminations in carbon fiber reinforced polymer composite CT using convolutional neural networks. In AIP Conference Proceedings; American Institute of Physics: College Park, MD, USA, 2016; Volume 1706, p. 110014. [Google Scholar]
- Abdeljaber, O.; Avci, O.; Kiranyaz, M.; Boashash, B.; Sodano, H.; Inman, D. 1-D CNNs for structural damage detection: Verification on a structural health monitoring benchmark data. Neurocomputing
**2018**, 275, 1308–1317. [Google Scholar] [CrossRef] - Jeng, S.; Huang, Y. Time series classification based on spectral analysis. Commun. Stat.-Simul. Comput.
**2007**, 37, 132–142. [Google Scholar] [CrossRef] - Yanez-Borjas, J.; Camarena-Martinez, D.; Valtierra-Rodriguez, M.; SaucedoDorantes, J.; Amezquita-Sanchez, J. Methodology based on statistical features and linear discriminant analysis for damage detection in a truss-type bridge. In Proceedings of the 2019 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC), Ixtapa, Mexico, 13–15 November 2019. [Google Scholar]
- Zang, C.; Imregun, M. Structural damage detection using artificial neural networks and measured FRF data reduced via principal component projection. J. Sound Vib.
**2001**, 242, 813–827. [Google Scholar] [CrossRef] - Wong, S.C.; Gatt, A.; Stamatescu, V.; McDonnell, M.D. Understanding data augmentation for classification: When to warp? In Proceedings of the 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA), Gold Coast, Australia, 30 November–2 December 2016; 2016. [Google Scholar]
- Filippatos, A.; Langkamp, A.; Kostka, P.; Koch, I.; Böhm, R.; Gude, M. Gradual damage behaviour of polar orthotropic glass-fibre reinforced epoxy rotors; experimental and simulation analysis. In Proceedings of the 18th European Conference on Composite Materials (ECCM18), Athens, Greece, 25–28 June 2018. [Google Scholar]
- Filippatos, A.; Langkamp, A.; Gude, M. Influence of Gradual Damage on the Structural Dynamic Behaviour of Composite Rotors: Simulation Assessment. Materials
**2018**, 11, 2453. [Google Scholar] [CrossRef][Green Version] - Filippatos, A.; Gude, M. Influence of Gradual Damage on the Structural Dynamic Behaviour of Composite Rotors: Experimental Investigations. Materials
**2018**, 11, 2421. [Google Scholar] [CrossRef][Green Version] - Cuntze, R.; Freund, A. The predictive capability of failure mode concept-based strength criteria for multidirectional laminates. Compos. Sci. Technol.
**2004**, 64, 343–377. [Google Scholar] [CrossRef] - Hornig, A.; Böhm, H.; Modler, N.; Gude, M. Novel Design Methods for Composite Structures under High-Strain-Rate Loading Conditions. J. Fail. Anal. Prev.
**2019**, 19, 144–146. [Google Scholar] [CrossRef] - Richter, J.; Wiegand, J.; Kuhtz, M.; Hornig, A.; Gude, M. Deformation and Failure of Multi-Layered Fibre-Metal-Laminates Subjected to Highly-Dynamic Loadings Conditions. In Proceedings of the 18th European Conference on Composite Materials (ECCM-18), Athens, Greece, 24–28 June 2018. [Google Scholar]
- Van Rossum, G.; Drake, F.L. Python 3 Reference Manual; CreateSpace: Scotts Valley, CA, USA, 2009. [Google Scholar]
- Izenman, A. Modern Multivariate Statistical Techniques: Regression, Classification and Manifold Learning; Springer: New York, NY, USA, 2008. [Google Scholar]
- Sorzano, C.O.S.; Vargas, J.; Montano, A.P. A survey of dimensionality reduction techniques. arXiv
**2014**, arXiv:1403.2877. [Google Scholar] - Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. Adv. Neural Inf. Process. Syst.
**2012**, 25, 1097–1105. [Google Scholar] [CrossRef] - Perez, L.; Wang, J. The effectiveness of data augmentation in image classification using deep learning. arXiv
**2017**, arXiv:1712.04621. [Google Scholar]

**Figure 1.**A schematic illustration of the rotor geometry is shown on the left. In the middle, the generated FE-mesh is presented with one element in the thickness direction and an approximate element thickness of 5 mm. On the right, the boundary conditions are illustrated.

**Figure 3.**The four physical-based classifiers and their labels. (

**a**) shows the radial position, (

**b**) the angular position and (

**c**) the load magnitude that describes the out-of-plane load. The out-of-plane load with damage accumulation is shown in (

**d**) [11].

**Figure 4.**Correlation matrices for 5000 frequencies from the spectrum. (

**a**) shows the Pearson correlation matrix for linear dependencies and (

**b**) the Spearman correlation matrix for nonlinear dependencies.

**Figure 5.**Architecture of NNs used for classification. (

**a**) shows the FC. Fully connected (dense) layers were used in combination with the rectified linear activation function (ReLU) and the softmax activation function in the output layer. Dropout layers were inserted to reduce overfitting. (

**b**) visualizes the structure of the CNN, with 1D-convolutional and pooling layers being fundamental to this network type.

**Figure 6.**Classification results of two different network architectures FC (

**a**) and CNN (

**b**). We compare the models FC-OD-Pure-RAD, FC-OD-Pure-ANG, FC-OD-Pure-MAGN and FC-OD-Pure-DAM (

**a**) with CNN-OD-Pure-RAD, CNN-OD-Pure-ANG, CNN-OD-Pure-MAGN and CNN-OD-Pure-DAM (

**b**). The abbreviations can be looked up in Appendix A. OD stands for original data.

**Figure 7.**Accuracy for each of the dimensionality reduction methods PCA, KPCA and using only every 30th frequency point depends on the number of components. Representative curves are shown for the classifier detecting the radial position. The encircled points indicate the chosen number of components.

**Figure 8.**Comparison of dimensionality reduction methods for the models FC-OD-30thFreq-RAD/-MAGN/-DAM, FC-OD-PCA-RAD/-MAGN/-DAM, FC-OD-PCA+LDA-RAD/-MAGN/-DAM, FC-OD-SVD+LDA-RAD/-MAGN/-DAM and FC-OD-KPCA-RAD/-MAGN/-DAM. Accuracy calculated with FC is shown. The dimensionality reduction methods are PCA, PCA+LDA, SVD+LDA, KPCA and using only every 30th data point. Angular position was omitted because there was no training effect at all.

**Figure 9.**Classification results with an augmented data set for the two network architectures: FC (

**a**) with models FC-AD-Pure-RAD/-ANG/-MAGN/-DAM and CNN (

**b**) with models CNN-AD-Pure-RAD/-ANG/-MAGN/-DAM. The data set contains in total 7920 spectra. Three bars corresponding to accuracy, sensitivity and specificity are shown.

Factor | Unit | Levels |
---|---|---|

impact load | kN | 8, 12, 16, 20 |

radius | mm | 75, 105, 135, 165, 195, 225 |

angle | ° | 0, 45, 90 |

centrifugal load | % | 51, 53, 56, 58, 59, 60, 63, 69, 76, 100 |

**Table 2.**Hyperparameters used for the networks FC and CNN depending on the four classifiers RAD, ANG, MAGN and DAM.

Parameter | FC-RAD | FC-MAGN | FC-DAM | CNN-RAD | CNN-DAM |
---|---|---|---|---|---|

FC-ANG | CNN-MAGN | ||||

CNN-ANG | |||||

optimizer | Adam | Adam | Adam | Adam | Adam |

learning rate | 0.0001 | 0.0001 | 0.0001 | 0.001 | 0.0001 |

epochs | 100 | 100 | 100 | 100 | 100 |

batch size | 64 | 64 | 64 | 64 | 16 |

filter | - | - | - | 40, 40, 20 | 90, 90, 60 |

kernel size | - | - | - | 9 | 3 |

pooling size | - | - | - | 6 | 3 |

dense layer | 1000, 100, 100 | 5010, 1000, 100 | 6000, 1000, 100 | 20 | 20 |

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |

© 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Scholz, V.; Winkler, P.; Hornig, A.; Gude, M.; Filippatos, A. Structural Damage Identification of Composite Rotors Based on Fully Connected Neural Networks and Convolutional Neural Networks. *Sensors* **2021**, *21*, 2005.
https://doi.org/10.3390/s21062005

**AMA Style**

Scholz V, Winkler P, Hornig A, Gude M, Filippatos A. Structural Damage Identification of Composite Rotors Based on Fully Connected Neural Networks and Convolutional Neural Networks. *Sensors*. 2021; 21(6):2005.
https://doi.org/10.3390/s21062005

**Chicago/Turabian Style**

Scholz, Veronika, Peter Winkler, Andreas Hornig, Maik Gude, and Angelos Filippatos. 2021. "Structural Damage Identification of Composite Rotors Based on Fully Connected Neural Networks and Convolutional Neural Networks" *Sensors* 21, no. 6: 2005.
https://doi.org/10.3390/s21062005