Treffer: Automating the Certificate Verification Process

Automation has seen a rapid growth during the recent decade and has evolved almost every industry, as it has allowed processes to become more reliable and efficient. One of the main objectives with automation is to reduce or eliminate time-consuming and tedious repetitive tasks to allow time to be allocated to more important tasks. Similarly, this thesis sought to explore how the process of manually verifying certificates at a case company specialized in calibration equipment manufacturing and services could be improved with modern tools such as machine learning to verify that the measurement results in the certificates were correct while simple rule-based approaches could be applied to other parts of the certificate where faults usually occurred to create an assistant to aid technicians during the verification process. To structure the thesis, a simplified version of the CRISP-DM framework was used, which consisted of four different phases. First, a focus group interview with the chief of the labora-tory and technicians was held to map out how the current process worked, what the data in the certificates implied, where faults usually occurred and what kind of solution was desired. These answers were used as the requirements during the development of a potential solu-tion. Second, methods to prepare the certificate data were developed, both to prepare data sufficient enough to train a model as well as being able to extract data from the certificate which had to be verified. The third phase consisted of developing the models, where seven different models were compared and evaluated, out of which four were selected for further evaluation. In the last phase, the performance of the selected models was evaluated where unseen data was used as the input and the prediction the model made as the output. The results indicated that the selected machine learning models all performed exceptionally well and were able to make accurate predictions, especially the Extra Trees algorithm showed promising results on the two ...