https://www.mdu.se/

Modern heat exchanger factories test every unit by recording a one-dimensional spatial sequence and relying on a human operator to spot anomalies, a slow and repetitive process that can generate false alarms or allow faulty exchangers to reach customers. This thesis investigates whether supervised machine learning can automate inspections and improve reliability, addressing two core challenges: extreme class imbalance and variable sequence lengths.For tree-based models, we engineered 149 statistical, spectral and shape features. For deep learning, we resampled each signal to 500 points and trained with focal loss. We compared XGBoost, a convolutional neural network and Transformer hybrid called ConvTran, and four ensemble strategies under stratified five-fold cross-validation.Focusing on the minority (defect) class, XGBoost achieved an F1 score of 73 percent, precision of 73 percent, and a recall of 76 percent. ConvTran detected more faults (recall at 90 percent) but produced many false alarms (precision at 38 percent). By blending their strengths, the weighted adaptive ensemble reached precision of 82 percent, recall of 90 percent and an F1 score of 86 percent. A model-switching ensemble increased precision to 89 percent with recall at 84 percent, achieving overall accuracy of 98.6 percent.These findings demonstrate that automated anomaly detection is feasible with further refinement. A hybrid workflow that flags low-confidence predictions for manual inspection while fully automating high-confidence predictions appears most promising.