Defective Pump Impeller Detection Using a CNN

This repository contains the files of my project for the "Artificial Intelligence" course at Amirkabir University of Technology (Tehran Polytechnic).

Project Description

This project deals with a classification problem. Since we are trying to classify images, we have decided to solve it with a convolutional neural network due to its abilities in image processing.

Dataset

The data we used are available on Kaggle.

Examples

Defective Pump Impeller

Non-Defective Pump Impeller

The Convolutional Neural Network

We created a CNN with the following architecture:

• Conv2D: 64x64x8
• MaxPool: 32x32x8
• Conv2D 32x32x8
• MaxPool: 16x16x8
• Flatten: 2048
• Dense: 16
• Dense: 16
• Dense: 1

Total params: 33,737

Training and Validation

Accuracy

We used binary_accuracy as the performance metric of the network.

Loss

We used binary_crossentropy as the loss function of the network.

Results

Our CNN model can classify images of pump impeller with an accuracy of ~96%.