
SWIFT
Sensor-based Wind turbine bearing Integrity and Fault Tracking
Facts
Coordinator: VUB — Prof. Jan Helsen
Partners: VUB and 3E
Project type: R&D&I — Fundamental and Industrial Research
Start date: November 2025
Duration: 3 years (36 months)
Budget: €1,096,165
Project context
Belgium’s offshore wind farms have matured rapidly, becoming a vital pillar of the national energy transition. However, operation and maintenance (O&M) activities remain costly and depend heavily on turbine manufacturers (OEMs) for data access and diagnostic services. This dependency limits the ability of local asset owners and service providers to perform predictive maintenance or develop their own digital tools. In many cases, operators have limited access to the high-frequency measurement data that is required to assess component health.
SWIFT addresses this challenge by enabling independent, sensor-based monitoring and fault tracking of wind turbine bearings, one of the most critical components in drivetrain reliability. The project will empower Belgian operators to perform advanced condition monitoring using affordable hardware and in-house analytics.
Project description
SWIFT develops and validates a sensor-based, data-driven monitoring framework for wind turbine main bearings and gearbox bearings. The project combines edge sensing, vibration analysis, and artificial intelligence to detect and predict bearing faults even when full OEM data or design details are unavailable.
The work is structured in five technical work packages:
WP1 – Low-cost sensor technology and data acquisition: design and validation of an add-on sensing setup capable of capturing relevant vibration and operational data from turbines.
WP2 – Failure annotation and reference database: creation of a standardized dataset linking field observations, vibration signatures, and failure cases.
WP3 – Blind vibration-based diagnostics: development of analysis methods that can identify bearing faults without requiring confidential OEM models or geometry.
WP4 – Fleet-level SCADA anomaly detection: extraction of health indicators and anomaly patterns from existing SCADA data across multiple turbines and sites.
WP5 – Data fusion and decision support: integration of the developed algorithms into a hybrid framework that combines vibration and SCADA insights to generate actionable maintenance recommendations.
The SWIFT methodology will be validated using data from the Norther offshore wind farm, serving as an industrial case study.
Project objectives
Develop a low-cost and independent sensing solution for drivetrain bearings.
Establish a failure-annotation framework and open reference database.
Create OEM-independent diagnostic tools for vibration data.
Build AI-based models for SCADA anomaly detection at fleet level.
Combine vibration and SCADA indicators in an integrated fault tracking platform.
Deliver automated diagnostic and maintenance recommendations to operators.
Demonstrate quantitative impact:
False positives reduced by 60%
O&M costs reduced by 8%
Lost production reduced by 11%
Impact in Belgium
SWIFT will strengthen the Belgian offshore wind ecosystem by reducing reliance on OEMs and enabling local players to perform advanced asset monitoring. By improving failure prediction and optimizing maintenance planning, the project directly contributes to lowering the Levelized Cost of Energy (LCOE) for offshore wind. It also supports the creation of new digital service offerings by Belgian SMEs and strengthens academic–industrial collaboration in smart maintenance technologies.
Dissemination of results
Project outputs will be made available through public reports, datasets, and open-access publications.
Toolboxes / Datasets
Sensor configuration guidelines and anonymized vibration datasets
Open scripts for SCADA-based anomaly detection
Reports
Annual technical progress reports (public summaries)
Final technical report with key results
Publications
JOURNAL ARTICLES:
Interpretable U-Net RCNN for blind peak detection in vibration spectra
G. Protopapadakis, P. Borghesani, C. Peeters, N. Herwig, W. A. Smith, Z. Peng, J. Helsen
Mechanical Systems and Signal Processing, 260, 114857 (2026) - Read the article
Summary: Finding the peaks in a vibration spectrum is the first step in most fault diagnosis, but noise and unknown spectral content make it hard to automate. This paper combines two deep-learning architectures (U-Net and region-based CNN) into a model that detects and locates spectral peaks without any prior knowledge of the signal or the machine. It was trained on realistic simulated signals, benchmarked against two established peak-identification methods, and tested on vibration data from offshore wind turbines.
Relevance to SWIFT: Blind peak detection is the foundation of SWIFT’s OEM-independent diagnostics (WP3). It lets the project infer gear mesh frequencies and fault sidebands directly from measured vibrations, without the confidential drivetrain design data that operators usually cannot obtain.
Condition monitoring of wind turbine drivetrains: state-of-the-art technologies, recent trends, and future outlook
K. Kestel, X. Chesterman, D. Zappalá, S. Watson, M. Li, E. Hart, et al., J. Helsen
Wind Energy Science, 11, 2103–2155 (2026) · Open access - Read the article
Summary: An international review of condition monitoring and prognostics for wind turbine drivetrains, among the most failure-prone and maintenance-intensive parts of a turbine. It covers SCADA-based methods, high-frequency vibration and acoustic analysis, machine learning and digital twins, and identifies the practical barriers that slow industrial adoption, such as limited data access, scarce fault data and uncertainty.
Relevance to SWIFT: The review sets out the state of the art that SWIFT builds on. It also documents the main problem the project addresses: operators often lack access to the data and design information needed for advanced monitoring.
CONFERENCE PAPERS:
Early fault detection in rotating machinery via multivariate autoencoder-based indicator fusion
F. Jamil, N. Sudhakaran, X. Liu, M. Stammler, A. Abrahamsen, C. Peeters, J. Helsen
PHM Society European Conference, 9(1), 1–10 (July 2026) · Open access - Read the paper
Summary: A deep-learning model learns the normal behaviour of a machine from healthy data only and merges many condition indicators into a single health indicator. It was run in real time on run-to-failure bearing experiments on a large bearing test rig. The model flagged faults at an earlier damage stage than conventional condition monitoring, and this was confirmed by inspecting the bearings after the tests.
Relevance to SWIFT: This is the normal-behaviour modelling approach used in SWIFT (WP1) to turn continuous vibration data into one interpretable health indicator per drivetrain. The test-rig validation, with physical inspection of the damage, shows that it detects bearing faults early, which is the core aim of SWIFT.
Conferences & Workshops
Presentations at EERA DeepWind, WindEurope, and Belgian Offshore Days
Final public workshop hosted by OWI-Lab at VUB
Project funded by




