Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This thesis aims to use the measurements provided by a novel sensor developed by iSensPro, to combat the problem of corrosion under insulation. More specifically, the problem of dry-wet classification is tackled using data driven machine learning models. The cost of gathering data samples is in practice very dependent on the parameters of the tests like the scale of the setup and the type of wetting, such as leakages or rain intrusion. Therefore, this study aims to train a classifier using only labeled samples of relatively cheap to perform tests while still achieving high performance on situations for which it is relatively expensive to gather labeled data. In this thesis, the situation where it is relatively cheap to gather labeled data is chosen to be the situation with moisture intrusion due to rain. The situation where it is relatively expensive to gather labeled data is chosen to be the situation with moisture intrusion due to a leakage. To perform this information transfer from one situation to another, transfer learning techniques are used. More specifically, several models using adversarial domain adaptation are implemented and tested. These models are trained on the labeled data from the situation with moisture intrusion due to rain while using unlabeled samples from the situation with a leakage. For the experimental design the reverse validation method is used to select the models' hyperparameters. The reliability of the reverse validation method is also examined. It is shown that the final selected transfer learning model is able to correctly classify all the dry samples of the leakage tests and in most of the tests the classifier is also able to detect all the wet samples due to leakage. However, in some of the performed experimental tests the model is not able to recognize any wet sample with moisture intrusion due to leakage.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The goal of this work is to investigate traditional machine learning and domain adaptation techniques for robust classification in scarce data settings, specifically targeting wind noise detection during End-of-Line testing in the automotive industry. Utilizing a synthetically generated dataset, derived from Virtual Sound Synthesis software and actual test recordings, the research focuses on the Power Spectral Density (PSD) of acoustic data as the feature representation. The Welch method was employed for PSD computation.The initial experiment explored varying frequency resolutions of PSD and their impact on the performance of machine learning pipelines comprising Principal Component Analysis (PCA) and classifiers such as Support Vector Machines (SVM) and Random Forests (RF). It was observed that high-frequency resolutions (nperseg > 16384) often led to overfitting, while values between 512 and 16384 were optimal, achieving high classification accuracy. The SVM classifier notably performed well for Mondeo and Focus datasets under specific hyperparameter configurations. Feature-based Domain Adaptation (DA) methods were compared in the next experiment, such as Subspace Alignment (SA), CORAL, Transfer Component Analysis (TCA), and kernel PCA, across multiple classifiers and scaling techniques. SA, combined with standard scaling and RF, consistently demonstrated superior performance, achieving over 85\% accuracy with more than 10 components in the subspace projected using SA. The study found that increased data availability only slightly improved performance and the performance was promising despite significantly lower resource settings, thus inciting further investigation. Overall, the RF classifier with 100 estimators consistently exhibited the highest classification performance, achieving over 98\% accuracy in initial experiments and 80\% in domain adaptation experiments, without significant degradation in scarce data settings. It was also seen that the SVM classifier had exhibited a Robust performance across the domains tested with 80\% accuracy using Subspace Alignment and 20 components in the scarcest data settings used. This research has demonstrated that SA, when combined with standard scaling, RF classifiers or SVM classifiers, is promising for transfer learning.