Electronic Theses and Dissertation

Pengecoran logam merupakan salah satu proses manufaktur yang banyak digunakan di berbagai sektor industri, namun masih rentan terhadap cacat hot tearing (retakan panas) yang dapat menurunkan kualitas produk. proses inspeksi secara manual memiliki keterbatasan dari sisi akurasi, konsistensi, dan efisiensi waktu. penelitian ini bertujuan mengembangkan sistem cerdas berbasis deep learning untuk mendeteksi dan mengklasifikasikan tingkat keparahan hot tearing menggunakan arsitektur cnn custom residual. dataset diperoleh melalui akuisisi citra spesimen hasil pengecoran logam di laboratorium fakultas teknik universitas syiah kuala. dataset awal terdiri atas 350 citra dan ditingkatkan menjadi 3.500 citra melalui augmentasi data. seluruh citra diklasifikasikan ke dalam lima kategori, yaitu no tear, hairline, moderate tearing, severe tearing, dan fully broken. arsitektur yang diusulkan dikembangkan dari cnn baseline ojha et al. (2025) melalui penambahan residual block, batch normalization, dan global average pooling. evaluasi dilakukan menggunakan metrik accuracy, precision, recall, dan f1-score, serta dibandingkan dengan mobilenetv2, resnet50, dan cracknet multiscale. cnn baseline memperoleh akurasi sebesar 88,76%, sedangkan cnn custom residual meningkatkan akurasi menjadi 97,14%. pada perbandingan model, mobilenetv2 memperoleh akurasi tertinggi sebesar 97,52%, diikuti cnn custom residual (97,14%), resnet50 (96,76%), dan cracknet multiscale (61,14%). selain itu, cnn custom residual menunjukkan efisiensi inferensi yang lebih baik dibandingkan mobilenetv2 pada pengujian yang dilakukan. model kemudian berhasil diimplementasikan ke dalam prototipe aplikasi android menggunakan flutter dan tensorflow lite. hasil penelitian menunjukkan bahwa pendekatan cnn berbasis residual learning mampu meningkatkan performa klasifikasi tingkat keparahan hot tearing serta berpotensi mendukung pengembangan sistem inspeksi kualitas pada industri pengecoran logam. kata kunci: deep learning, cnn custom residual, hot tearing, klasifikasi citra, pengecoran logam.

Metal casting is one of the most widely used manufacturing processes across various industrial sectors; however, it remains susceptible to Hot Tearing defects, which can significantly reduce product quality. Manual inspection methods have limitations in terms of accuracy, consistency, and time efficiency. This study aims to develop a Deep Learning-based intelligent system for detecting and classifying the severity levels of Hot Tearing using a Custom Residual Convolutional Neural Network (CNN) architecture. The dataset was acquired from images of cast metal specimens collected at the Laboratory of the Faculty of Engineering, Universitas Syiah Kuala. The initial dataset consisted of 350 images and was expanded to 3,500 images through data augmentation. All images were categorized into five severity levels: No Tear, Hairline, Moderate Tearing, Severe Tearing, and Fully Broken. The proposed architecture was developed from the CNN baseline introduced by Ojha et al. (2025) by incorporating Residual Blocks, Batch Normalization, and Global Average Pooling. Model performance was evaluated using accuracy, precision, recall, and F1-score, and compared with MobileNetV2, ResNet50, and CrackNet Multiscale. The CNN Baseline achieved an accuracy of 88.76%, while the proposed CNN Custom Residual improved the accuracy to 97.14%. In the comparative evaluation, MobileNetV2 achieved the highest accuracy (97.52%), followed by CNN Custom Residual (97.14%), ResNet50 (96.76%), and CrackNet Multiscale (61.14%). Furthermore, the proposed CNN Custom Residual demonstrated better inference efficiency than MobileNetV2 under the experimental conditions. The proposed model was successfully implemented in an Android prototype application using Flutter and TensorFlow Lite. The results indicate that the proposed Residual Learning-based CNN effectively improves the classification performance of Hot Tearing severity levels and has the potential to support automated quality inspection systems in the metal casting industry. Keywords: Deep Learning, Custom Residual CNN, Hot Tearing, Image Classification, Metal Casting