MODELING PISTON DAMAGE DETECTION USING A CONVOLUTIONAL NEURAL NETWORK BASED ON DIGITAL IMAGE

Utami, Lega Putri
Leni, Desmarita

International Journal of Multidisciplinary Research and Literature (IJOMRAL)
Vol. 3 No. 2 (2024): INTERNATIONAL JOURNAL OF MULTIDISCIPLINARY RESEARCH AND LITERATURE

Product inspection is a crucial component of product quality control, aiming to evaluate and ensure that products meet predefined standards. In this research, the modelling of piston damage detection is conducted using a Convolutional Neural Network (CNN). The dataset employed consists of images of pistons categorized into three groups: Defected1, Defected2, and Normal. Two hundred eighty-five images are utilized as training data, with the data distribution percentages for Defected1, Defected2, and Normal being 30.9%, 34.4%, and 34.7%, respectively. The model is validated using newly generated data through augmentation techniques, resulting in 60 images. The CNN model uses a sequential Keras architecture comprising convolutional layers, pooling layers, fully connected layers, and softmax activation. The Adam optimizer with a learning rate 0.0001 is employed for model training, with validation using a 5-fold cross-validation. The model is evaluated using the Loss, Accuracy, and Confusion Matrix, achieving a training accuracy of 0.722 and a validation accuracy of 0.689. An early stopping function is applied to halt training when there is no improvement in validation accuracy. The confusion matrix results indicate that the model adequately classifies data with Accuracy, Recall, and Precision values of 69%, 69%, and 70%, respectively