Defect-detection in casting products using grid search augmented PCA CNN hybrid model

<b>Abstract. </b>Automated and accurate defect detection techniques are increasingly becoming more important in manufacturing systems. These techniques are commonly mentioned in academic literature and implemented by industry. Convolutional Neural Networks (CNNs) are popular tools used in image recognition allowing them to be used for defect detection systems in manufacturing. Principal component analysis (PCA) is also an important statistical technique used for feature extraction. This study proposes a grid search augmented PCA-CNN hybrid method for defect-detection in casting products. Proposed approach extracts the most significant features of casting product image data set by selecting the top principal components which leads to dimensionality reduction in data. This in turn leads to retention of the most informative aspects and a reduction in noise improving the efficiency of algorithms. Preprocessing done by applying PCA reduces the computation time of algorithms and improves accuracy when training data for CNNs is limited. Grid search was also implemented into the model to select the best model parameters settings based on model accuracy and loss. We aim to achieve enhanced performance of real-time monitoring and defect-detection in manufacturing systems by integrating these techniques. Proposed hybrid model’s results are compared with a standard CNN model and improved performance in nearly all model parameters is observed. However proposed approach has its challenges such as the quality of input data, complexity and training of the CNN model, selection of features can hinder the performance of the proposed system. The sparsity of training data set is also an important challenge as it leads to model overfitting and reduction in models generalization capabilities. Therefore, there is an ongoing need for further research in this topic and exploration of new solutions such as transfer learning and multi-modal data fusion to improve defect detection.