Session: 41-03 AI and advanced methods for wind energy II

Submission Number: 177783

State-Space Based NBM for Anomaly Detection in Wind Turbines 

The worldwide spread of wind turbine power plants lead to the need for advanced monitoring and maintenance strategies and the implementation of fault detection strategies is a key step of this process, ensuring optimal performance and reduced downtimes. Supervisory Control and Data Acquisition (SCADA) systems play an indispensable role in wind energy production, as the data they collect serve as the sole real-time interface between the turbines and the operator. These data enable the proactive planning of maintenance interventions, ensuring operational efficiency and extending the lifespan of the assets. Despite traditional model-based approaches, data driven methods can learn directly from observed data, which means they can adapt to changes in the underlying distribution or system behavior over time. In particular machine learning and deep learning approaches effectively capture complexity and non-linear dynamics of wind turbine systems. The proposed methodology relies on a reconstruction-based anomaly detection framework. The key idea is to learn the turbine’s healthy conditions defining a Normal Behavior Mode (NBM) by training machine learning architectures exclusively on fault free SCADA. Once trained, the model reconstructs the input data of multivariate time series, and anomalies are detected when the reconstruction error exceeds a data-driven threshold. To effectively capture the complex temporal dependencies and dynamic patterns typical of SCADA data, the proposed approach adopts the Mamba architecture, a selective state-space model (SSM) specifically designed for efficient long-sequence modeling. Mamba conditions its internal parameters on the input, enabling it to selectively propagate or forget information while leveraging a hardware-efficient recurrent algorithm that ensures linear-time scalability. This architecture is selected to ensure the model accurately reconstructs normal operating dynamics while remaining computationally efficient over extended SCADA records. The study relies on a SCADA dataset provided by EDP (Energias de Portugal). The dataset spans two years of decaminute-resolution operational records collected from four wind turbines. It includes measurement variables provided by SCADA sensors and supplementary files documenting failure logs and status logs. The results highlight the capability of the Mamba-based SSM to effectively detect anomalies in SCADA data, exhibiting strong temporal consistency with recorded alarms and actual failure events. The proposed model is compared with state of art through widely adopted metrics indicating accuracy performance. In addition, Cumulative Sum (CUSUM) control charts are employed to further investigate the temporal characteristics of the detected anomalies, providing insights into their persistence and their progression prior to failure events.

Presenting Author: Gianmarco Bianchi Sapienza University of Rome

Presenting Author Biography: I am Gianmarco Bianchi, a first-year PhD student in Industrial and Management Engineering at Sapienza University of Rome, where I also earned my master degree in Mechanical Engineering. My research focuses on wind energy, specifically on data driven methods for power forecasting and anomaly detection.

Authors:

Valerio Francesco Barnabei Sapienza University of Rome
Gianmarco Bianchi Sapienza University of Rome
Gabriele Abbadessa Sapienza University of Rome
Tullio Carlo Maria Ancora Sapienza University of Rome
Alessandro Corsini Sapienza University of Rome