Investigation of the Capabilities of Artificial Intelligence in Predictive Maintenance of Technical Equipment
Predictive maintenance promises to curb unplanned downtime and extend asset life, yet real-world deployments struggle with scarce fault labels, strict privacy rules, and limited edge-compute budgets. This paper presents an end-to-end framework that combines a physics-based digital twin, contrastive self-supervised representation learning, and FedProx-style federated optimisation to deliver sub-second anomaly detection on low-power industrial edge devices while keeping raw data on-site. A 12 000-hour pilot on five induction-motor drives achieved an F1-score of 0.92—13 percentage points above the best fully supervised baseline—and issued a single true bearing-fault alert 19 days in advance, avoiding an estimated 32 500 Euro in downtime. The encoder–classifier stack fits in 18 MB of memory, runs at 2.4 ms latency on an NVIDIA Jetson Orin Nano, and exchanges only 6.5 MB of encrypted weight updates every six hours. These results indicate that self-supervision, synthetic replay, and privacy-preserving federation together form a scalable path toward industry-grade, regulation-compliant predictive maintenance.
Paper
Full text
Investigation of the Capabilities of Artificial Intelligence in Predictive Maintenance of Technical Equipment
Semantic Scholar · 2025
Abstract
Predictive maintenance promises to curb unplanned downtime and extend asset life, yet real-world deployments struggle with scarce fault labels, strict privacy rules, and limited edge-compute budgets. This paper presents an end-to-end framework that combines a physics-based digital twin, contrastive self-supervised representation learning, and FedProx-style federated optimisation to deliver sub-second anomaly detection on low-power industrial edge devices while keeping raw data on-site. A 12 000-hour pilot on five induction-motor drives achieved an F1-score of 0.92—13 percentage points above the best fully supervised baseline—and issued a single true bearing-fault alert 19 days in advance, avoiding an estimated 32 500 Euro in downtime. The encoder–classifier stack fits in 18 MB of memory, runs at 2.4 ms latency on an NVIDIA Jetson Orin Nano, and exchanges only 6.5 MB of encrypted weight updates every six hours. These results indicate that self-supervision, synthetic replay, and privacy-preserving federation together form a scalable path toward industry-grade, regulation-compliant predictive maintenance.