Project name: From Signals to Insight: Predicting Failures in Industrial Systems using AI
Project’s period: May 2024 – October 2024
Partner: PredictiveDataScience
The collaboration with PredictiveDataScience within the Hopero project focused on detecting changes in measured signals that indicate an upcoming fault in industrial devices, particularly electric motors. The goal was to move beyond basic signal monitoring towards a data-driven solution enabling reliable detection of anomalies, reduced downtime, and higher operational reliability. At the same time, the ambition was broader than vibrodiagnostics, focusing on continuous monitoring, advanced data analysis, and building a foundation for predictive maintenance across industrial applications.
HOW WE APPROACHED IT
PredictiveDataScience, in collaboration with KInIt designed and validated an anomaly detection approach based on advanced analytics and AI. Multiple methods were evaluated and selected according to performance and robustness in real industrial conditions. We addressed key challenges such as irregular sampling, missing data, and noisy signals. Strong emphasis was placed on validation, interpretability, and trustworthiness to ensure usability in practice. Our approach reflects a wider focus on monitoring, analysis, and prediction, not limited to vibration data. The applied AI methods, assessed in 2025 as trustworthy, support reliable deployment in industrial environments and enable future development towards predictive maintenance.
“Predicting an upcoming fault is inherently a difficult problem. It requires not only the ability to automatically identify previously unseen failure modes, but also to operate without labeled datasets, as is usually the case in unsupervised learning scenarios.”
MAREK LÓDERER
AI Specialist in Enviro team
WHAT WE DELIVERED
The solution was validated in an industrial setting and proved capable of reliably detecting anomalous motor behaviour while significantly reducing false alarms, with zero false anomalies and zero false predictive failures reported in production conditions. The outputs were confirmed by domain experts, ensuring both technical accuracy and practical relevance.
Beyond anomaly detection, the work delivered personalised time-to-failure predictions and personalised predictive maintenance recommendations tailored to individual equipment behaviour. As a direct result, OEE (Overall Equipment Effectiveness) improved by 15% on the monitored production line. In addition, the solution established a foundation for further scaling of AI-driven monitoring across different industrial domains.
“We would like to thank the KInIT team for their persistence, patience, and technical expertise, which helped us properly evaluate multiple anomaly detection approaches and address key challenges related to data discontinuities and irregular sampling. Their thorough validation of our detection outputs and their insights into future data representation for fault classification have given us confidence in our solution and provided a solid foundation for its practical deployment in industrial environments.”
