Modern industrial environments collect massive amounts of alarm and log data from their machinery and equipment. These data are fine-grained, detailed and highly dimensional, and in many real-world settings they are entirely unlabeled: it is not possible to know which observations are anomalies, which makes supervised algorithms impractical. Detecting anomalies and understanding their nature is nevertheless crucial to preventing system failures and optimizing equipment performance. The thesis addresses this problem in a real-world setting in collaboration with Zamperla, a global leader in the amusement ride industry that provided the dataset, and with Statwolf, a data analytics company. The analysis was conducted using their platform, which made it possible to handle the large volume of data generated by the ride. The objective was to create an unsupervised machine learning pipeline for a high-frequency dataset from an amusement ride known as Power Zurge. Initial steps involved dimensionality reduction via Principal Component Analysis (PCA) and the identification of anomalous days using clustering methods—such as HDBSCAN—or anomaly scoring algorithms like Isolation Forest. Subsequently, the anomalous days were analyzed to extract co-occurrence patterns using methods such as n-grams or the FP-growth algorithm, which generated association rules among alarms. Some of the methods and algorithms employed proved effective at identifying anomalous days and uncovering recurring patterns in the sets of alarm occurrences—without relying on dataset labels—thereby demonstrating their applicability in a real-world context. Keywords: Anomaly Detection, Unsupervised Machine Learning, Dimensionality Reduction, PCA, HDBSCAN, Isolation Forest, Co-occurrence Analysis, Association Rules, Alarm

Modern industrial environments collect massive amounts of alarm and log data from their machinery and equipment. These data are fine-grained, detailed and highly dimensional, and in many real-world settings they are entirely unlabeled: it is not possible to know which observations are anomalies, which makes supervised algorithms impractical. Detecting anomalies and understanding their nature is nevertheless crucial to preventing system failures and optimizing equipment performance. The thesis addresses this problem in a real-world setting in collaboration with Zamperla, a global leader in the amusement ride industry that provided the dataset, and with Statwolf, a data analytics company. The analysis was conducted using their platform, which made it possible to handle the large volume of data generated by the ride. The objective was to create an unsupervised machine learning pipeline for a high-frequency dataset from an amusement ride known as Power Zurge. Initial steps involved dimensionality reduction via Principal Component Analysis (PCA) and the identification of anomalous days using clustering methods—such as HDBSCAN—or anomaly scoring algorithms like Isolation Forest. Subsequently, the anomalous days were analyzed to extract co-occurrence patterns using methods such as n-grams or the FP-growth algorithm, which generated association rules among alarms. Some of the methods and algorithms employed proved effective at identifying anomalous days and uncovering recurring patterns in the sets of alarm occurrences—without relying on dataset labels—thereby demonstrating their applicability in a real-world context. Keywords: Anomaly Detection, Unsupervised Machine Learning, Dimensionality Reduction, PCA, HDBSCAN, Isolation Forest, Co-occurrence Analysis, Association Rules, Alarm

Unsupervised Learning for Anomaly Detection in IoT-Based Amusement Ride Monitoring

SCAPIN, ISMAEL
2025/2026

Abstract

Modern industrial environments collect massive amounts of alarm and log data from their machinery and equipment. These data are fine-grained, detailed and highly dimensional, and in many real-world settings they are entirely unlabeled: it is not possible to know which observations are anomalies, which makes supervised algorithms impractical. Detecting anomalies and understanding their nature is nevertheless crucial to preventing system failures and optimizing equipment performance. The thesis addresses this problem in a real-world setting in collaboration with Zamperla, a global leader in the amusement ride industry that provided the dataset, and with Statwolf, a data analytics company. The analysis was conducted using their platform, which made it possible to handle the large volume of data generated by the ride. The objective was to create an unsupervised machine learning pipeline for a high-frequency dataset from an amusement ride known as Power Zurge. Initial steps involved dimensionality reduction via Principal Component Analysis (PCA) and the identification of anomalous days using clustering methods—such as HDBSCAN—or anomaly scoring algorithms like Isolation Forest. Subsequently, the anomalous days were analyzed to extract co-occurrence patterns using methods such as n-grams or the FP-growth algorithm, which generated association rules among alarms. Some of the methods and algorithms employed proved effective at identifying anomalous days and uncovering recurring patterns in the sets of alarm occurrences—without relying on dataset labels—thereby demonstrating their applicability in a real-world context. Keywords: Anomaly Detection, Unsupervised Machine Learning, Dimensionality Reduction, PCA, HDBSCAN, Isolation Forest, Co-occurrence Analysis, Association Rules, Alarm
2025
Unsupervised Learning for Anomaly Detection in IoT-Based Amusement Ride Monitoring
Modern industrial environments collect massive amounts of alarm and log data from their machinery and equipment. These data are fine-grained, detailed and highly dimensional, and in many real-world settings they are entirely unlabeled: it is not possible to know which observations are anomalies, which makes supervised algorithms impractical. Detecting anomalies and understanding their nature is nevertheless crucial to preventing system failures and optimizing equipment performance. The thesis addresses this problem in a real-world setting in collaboration with Zamperla, a global leader in the amusement ride industry that provided the dataset, and with Statwolf, a data analytics company. The analysis was conducted using their platform, which made it possible to handle the large volume of data generated by the ride. The objective was to create an unsupervised machine learning pipeline for a high-frequency dataset from an amusement ride known as Power Zurge. Initial steps involved dimensionality reduction via Principal Component Analysis (PCA) and the identification of anomalous days using clustering methods—such as HDBSCAN—or anomaly scoring algorithms like Isolation Forest. Subsequently, the anomalous days were analyzed to extract co-occurrence patterns using methods such as n-grams or the FP-growth algorithm, which generated association rules among alarms. Some of the methods and algorithms employed proved effective at identifying anomalous days and uncovering recurring patterns in the sets of alarm occurrences—without relying on dataset labels—thereby demonstrating their applicability in a real-world context. Keywords: Anomaly Detection, Unsupervised Machine Learning, Dimensionality Reduction, PCA, HDBSCAN, Isolation Forest, Co-occurrence Analysis, Association Rules, Alarm
Anomaly Detection
Unsupervised Learn.
Machine Learning
Industrial IoT
File in questo prodotto:
File Dimensione Formato  
Scapin_Ismael.pdf

Accesso riservato

Dimensione 2.54 MB
Formato Adobe PDF
2.54 MB Adobe PDF

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/119114