Undergraduate Research Work

Chittagong University of Engineering & Technology (CUET)

Thesis Title: Detection of Knocking in SI Engine Using a Deep Learning Approach: A Method Based on Convolutional Neural Network

Thesis Supervisor: Dr. Kazi Afzalur Rahman, Professor, Department of Mechanical Engineering, CUET.

Thesis Description: A deep learning model was trained to detect engine knocks automatically using ResNet-34, ResNet-18, and ResNet-50 architectures. A set of engine sounds was collected from different surrounding sources and audio spectrograms were generated to prepare a large dataset. It has been used to train and evaluate the model. Several audio processing applications have demonstrated promising results with convolutional neural networks (CNNs), which have been used in the approach. In order to prevent overfitting, the CNNs contained multiple convolutional layers with different kernel sizes and filter numbers, followed by max-pooling layers and dropout layers. Then the best resulting and trained model was exposed using a Web Application where anyone could upload a soundtrack of an engine. In return, the web application would give him/her the result if the engine was knocking.