شناسايي مكان و اندازه تورق در يك چندلايه متعامد كامپوزيتي مستطيلي با استفاده از يادگيري ماشين كلاسيك و شبكه‌هاي عصبي عميق

Identification of the location an‎d size of delamination in rectangular cross-ply composite laminate using classical machine learning an‎d deep neural networks

مهندسي مكانيك

Delamination detection in composite laminates is a topic of interest to many researchers. In this study, delamination detection in composite laminates was carried out using vibration methods an‎d machine learning. Initially, modal analysis was performed for intact an‎d delaminated composite plates using the finite element method an‎d natural frequencies were obtained for training the neural network. The results were then compared with experimental data for validation. Next, a novel deep ensemble neural network was designed an‎d the location an‎d size of the delamination were determined. This neural network integrated a multilayer perceptron (MLP), boosting trees, bagging trees, an‎d a gated recurrent unit (GRU), which were sequentially combined an‎d trained in multiple stages for delamination identification. Hyperparameters are optimized via grid search an‎d cross-validation method, an‎d a sensitivity analysis is performed to eva‎luate the influence of each natural frequency on the model. In the next stage, classical machine learning methods were employed for delamination detection. These methods include linear regression, support vector machines, Gaussian process regression, an‎d tree-based methods. The support vector machine methods used in this study include quadratic, cubic, an‎d Gaussian kernel models. Also, Gaussian process regression models with exponential an‎d Matérn 2.5 kernels were used to predict the location an‎d size of the delamination. At each stage, the root mean square error (RMSE) an‎d coefficient of determination (R-squared) were calculated to eva‎luate model performance an‎d accuracy. The results show that the designed ensemble network achieved lower error compared to classical machine learning methods. Among classical approaches, linear regression exhibited the highest error, while Gaussian process regression had the lowest error for detecting the location of the delamination. The best performance in predicting the size of the delamination was obtained using the quadratic support vector machine method. The innovation of this research lies in the design of an ensemble deep recurrent network through which delamination detection in composite laminates was performed with high accuracy.

6