With new technology, the automotive industry is progressing fast. The era of industry 4.0 and 5.0 has evolved the industry towards applied Artificial Intelligence. Aurobay in Skövde has realised the potential for AI solutions in visual quality inspection. This has the potential to improve quality assurance alongside aiding the operators with decision-making support. The visual quality control station located at the cylinder head manufacturing line visually inspects 100% of all the parts in the process manually. The purpose of this thesis is to investigate the possibility of a future implementation and present a proof of concept that serves as a foundation for a future real-world application where the tedious manual labour is removed from the station.
The project is based on Design Science Research where a six-step plan was utilised. Firstly, problem identification is undertaken to articulate the problem statement. Next, objectives are defined to navigate the course of the project through an extensive literature review. During the design and development phase, the artifact is created. Subsequently, in the demonstration phase, experiments are conducted. Evaluation involves analysing the results of the experiments. Finally, communication is the discussion of the thesis report.
Six experiments have been carried out on the created artifact, four is based on improving the baseline performance, while the fifth explores the amount of data needed to achieve satisfactory accuracy. The last experiment validates the functionality of simulated tiling spanning across different regions of the surface of the part. The first experiment conducted on full size images without abnormal directory yielded five correctly classified images and F1-score of 29%. The second experiment extended the first experiment by applying region of interest on the images leading to unchanged results. The third experiment evaluated the difference in performance between PaDIM and PatchCore using simulated tiling without abnormal directory. PaDIM managed to classify five correctly with a F1-score of 29%, while PatchCore managed to classify nine correctly with a F1-score of 32%. The fourth experiment aimed to compare the performance of PatchCore conducted in the previous experiment to the same model but with an abnormal directory. This resulted in a leap in performance, by adding an abnormal directory the model managed to classify 27 images correctly with a F1-score of 67%. The fifth experiment used the same model as the previous experiment, the only thing varying is the sizes of training-sets, in the previous experiment 100 images were used in the training-set, in addition to this result, the model was tested with 80, 60, 40, and 20 images respectively. While the number of correct classifications and F1-score remained the same. In addition to the tile that have been used in the previous experiments, four additional tiles were evaluated using the same model as in the previous experiment. The results remained almost unchanged if not accounting for the difference in the images, meaning that the simulated tiling works anywhere on the surface.
The findings and experience gained from this project is discussed. Recommendations and future work are proposed, as a next step in the process. The key steps are to analyse the camera location within the cell, exploring different approaches to present all characteristics for the vision system, and finally gather data to reduce time in the implementation phase.
2024. , p. viii, 64