Abstract

The goal of the research is the analysis of the usability of unsupervised anomaly detection (AD) methods for detection of valve plate failures in piston pumps. The unsupervised AD methods play a special role in predictive maintenance since they do not require failure data to be present during training, which represents a typical scenario of real-world manufacturing operations where faulty states are rarely recorded. A comprehensive comparative analysis of nine detection algorithms, including deep learning models, was conducted using real-world process data. The results assess the models' effectiveness and resilience to industrial noise, identifying optimal solutions for increasing the reliability of monitoring systems in the era of Industry 4.0. The best average result AUC=0.9150 was obtained by the AutoEncoder. The results show that the Isolation Forest model significantly differs from Local Outlier Factor and One-Class SVM in terms of feature importance. Four techniques stand out for their results.

Recommended Citation

Rojek, M. & Blachnik, M.(2026). Developing a Diagnostic Information System for Hydraulic Piston Pumps: A Comparative Analysis of Anomaly Detection Methods. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.59

Paper Type

Short Paper

DOI

10.62036/ISD.2026.59

Share

COinS
 

Developing a Diagnostic Information System for Hydraulic Piston Pumps: A Comparative Analysis of Anomaly Detection Methods

The goal of the research is the analysis of the usability of unsupervised anomaly detection (AD) methods for detection of valve plate failures in piston pumps. The unsupervised AD methods play a special role in predictive maintenance since they do not require failure data to be present during training, which represents a typical scenario of real-world manufacturing operations where faulty states are rarely recorded. A comprehensive comparative analysis of nine detection algorithms, including deep learning models, was conducted using real-world process data. The results assess the models' effectiveness and resilience to industrial noise, identifying optimal solutions for increasing the reliability of monitoring systems in the era of Industry 4.0. The best average result AUC=0.9150 was obtained by the AutoEncoder. The results show that the Isolation Forest model significantly differs from Local Outlier Factor and One-Class SVM in terms of feature importance. Four techniques stand out for their results.