Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Improving the performance of deep learning techniques using nature inspired algorithms and applying them in porosity prediction

Alamri, Nawaf ORCID: 2023. Improving the performance of deep learning techniques using nature inspired algorithms and applying them in porosity prediction. PhD Thesis, Cardiff University.
Item availability restricted.

[thumbnail of PhD Thesis]
PDF (PhD Thesis) - Accepted Post-Print Version
Download (4MB) | Preview
[thumbnail of Cardiff University Electronic Publication Form] PDF (Cardiff University Electronic Publication Form) - Supplemental Material
Restricted to Repository staff only

Download (119kB)


Within the field of Artificial Intelligence (AI), Deep Learning (DL) based on Convolutional Neural Network (CNN) can be used for analysing images. However, the performance of the DL models depends on the design of the CNN topology to achieve their best performance. Hence, firstly, this work addresses this problem by proposing a novel nature inspired hybrid algorithm called BA-CNN where a swarm based Bees Algorithm (BA) is used to optimize the CNN parameters. In addition, another algorithm called BABO-CNN is proposed that combines the BA with Bayesian Optimization (BO) to increase the CNN performance and that of BA-CNN and BO-CNN. This study shows that applying the hybrid BA-CNN to the ‘Cifar10DataDir’ benchmark image did not improve the validation and testing accuracy compared to the existing CNN and BO-CNN. However, the hybrid BA-BO-CNN achieved better validation accuracy of 82.22% compared to 80.34% and 80.72% for the CNN and BO-CNN, and also with a better testing accuracy of 80.74% compared to 80.54% and 80.69% for the CNN and BO-CNN respectively. The BA-BOCNN achieved lower computational time than the BO-CNN algorithm by 2 minutes and 11 seconds. Although applying both algorithms to the ‘digits’ dataset produced almost similar accuracies with a difference of 0.01% between BA-CNN and BO-CNN, the BA-CNN achieved a computational time reduction of 4 minutes and 14 seconds compared to the BOCNN, making it the best algorithm in terms of cost-effectiveness. Applying BA-CNN and BA-BO-CNN to identify ‘concrete cracks’ images produced almost similar results to some of the other existing algorithms with a difference of 0.02% between BA-CNN and original CNN. Finally, applying them to the ‘ECG’ images improved the testing accuracy from 90% for the BO-CNN to 92.50% for the BA-CNN and 95% for the BA-BO-CNN with a similar trend for validation accuracy and computational time. Secondly, the CNN that was adopted for the purpose of regression which is called RCNN was applied in the manufacturing context, particularly to predict the percent of porosity in the finished Selective Laser Melting (SLM) parts. Because testing the performance of the RCNN algorithm requires a large amount of experimental data which is generally difficult to obtain, in this study an artificial porosity image creation method is proposed where 3000 artificial porosity images were created mimicking real CT scan slices of the SLM part with a similarity index of 0.9976. Applying the RCNN to the 3000 artificial ii porosity images slices showed the porosity prediction accuracy to improve from 68.60% for the image binarization method to 75.50% for the RCNN, while the proposed novel hybrid BA-BO-RCNN and BA-RCNN yielded better prediction accuracies of 83% and 85.33% respectively. Thirdly, in order to improve the performance even further, this study proposes to add Long Short Term Memory (LSTM) to BA-CNN because of their ability to deal with sequential data to produce another novel hybrid algorithm called BA-CNN-LSTM and the results showed an increase in the prediction accuracy reaching 95.50%

Item Type: Thesis (PhD)
Date Type: Completion
Status: Unpublished
Schools: Engineering
Uncontrolled Keywords: Deep Learning Convolutional Neural Network Bees Algorithm Selective Laser Melting Artificial Porosity Images
Date of First Compliant Deposit: 18 May 2023
Last Modified: 18 May 2023 11:10

Actions (repository staff only)

Edit Item Edit Item


Downloads per month over past year

View more statistics