| dc.description.abstract | Anomaly detection in industrial time series data is essential for identifying and preventing potential issues in production processes, ensuring high product quality and reducing downtime. This master's thesis investigates the performance of two unsupervised machine learning algorithms, Local Outlier Factor (LOF) and DBSCAN, for detecting clogging events in the production process of Bioco, a company in the biotechnology industry. The main objective is to evaluate the algorithms' ability to provide early warnings for clogging events, enabling timely preventive actions.
The research process involves a thorough theoretical overview, data exploration and preprocessing, application of the selected algorithms, and evaluation of their performance. The study also examines the sensitivity of the algorithms to parameter tuning and the effectiveness of incorporating lagged variables as features in the anomaly detection models.
The results indicate that both LOF and DBSCAN can detect relevant anomalies in the time series data, but their performance in providing early warnings for clogging events is limited. While LOF requires careful parameter tuning, DBSCAN demonstrates more stable performance across different parameter settings. The inclusion of lagged variables does not improve the detection of clogging events, showcasing challenges in selecting the optimal lag length.
This study contributes to the existing literature on anomaly detection in industrial time series data by providing insights into the practical performance of LOF and DBSCAN algorithms in a specific industrial context. The findings highlight the importance of considering the effects of lagged variables and parameter tuning when designing anomaly detection models for industrial applications. Future research could explore other anomaly detection algorithms and their performance in different industrial settings to enhance the generalizability of the results. | |
