Outlier Detection in Wind Turbine Frequency Converters Using Long-Term Sensor Data

Nils Schwenzfeier, Markus Heikamp, Ole Meyer, Andre Hönnscheidt, Michael Steffes, Volker GruhnAAAI 2022 · DOI

Introduction

Operation and maintenance (O&M) costs represent a significant portion of a wind turbine's lifecycle expenses, often reaching up to 30%. At the heart of these machines is the frequency converter, a critical component that synchronizes generated energy with the power grid. In a Double-Fed Induction Generator (DFIG) setup, the converter manages the power flow between the rotor and the grid, making its reliability paramount. Failures in these converters lead to substantial downtime and high repair costs. Traditional maintenance strategies are typically reactive or preventive; however, there is a growing need for predictive maintenance to identify "silent failures" before they escalate into catastrophic breakdowns.

Methodology: LSTM Autoencoders with 1D-CNN

The researchers propose a self-supervised learning approach using an LSTM Autoencoder architecture enhanced with 1D convolutional layers. The model is designed to learn the normal operating patterns of the frequency converter by compressing input sensor data into a lower-dimensional bottleneck and subsequently reconstructing it.

The proposed LSTM Autoencoder architecture featuring 1D-CNN layers for feature extraction and LSTM layers for temporal sequence processing.

A key innovation in this architecture is the inclusion of 1D-CNN layers at the start of the encoder. These layers serve as a learnable preprocessing step, smoothing noisy sensor streams and extracting relevant features before they are processed by the LSTM units. This hybrid approach allows the model to capture both spatial correlations between different sensors and temporal dependencies over time. Without these convolutional filters, the high fluctuations in raw industrial IoT data often make standard LSTM models difficult to train effectively.

Data and Experimental Setup

The study utilized a 2.5-year dataset (August 2018 – December 2020) from five onshore wind turbines located in the same farm. Key monitored variables included Generator Rotor Speed, Mains Power (Active and Reactive), and DC Link Voltage. Data was resampled to 5-second intervals and segmented into 15-minute windows, resulting in 180 data points per window. All sensor values were normalized to a [0, 1] range to prevent bias toward sensors with higher absolute magnitudes.

To validate the model's sensitivity, the authors introduced synthetic outliers into the dataset. These "artificial outliers" simulated common failure modes, such as sensor value swaps, short circuits, and instances where rotor speed dropped to zero during active power generation.

Results and Expert Validation

The model's performance is measured by the reconstruction error. Since the autoencoder is trained primarily on normal operating data, it excels at reconstructing standard patterns but fails when encountering anomalies, resulting in a high error score that serves as an anomaly indicator.

Comparison of original normalized sensor data (blue) and model reconstructions (red dashed). High reconstruction errors indicate potential anomalies.

In synthetic tests, the model achieved a 1.0 recall for the top-50 highest error windows across all outlier types, perfectly separating alienated data points from the normal set. Real-world validation involved technical experts reviewing the windows with the highest reconstruction errors. These experts confirmed that the model successfully flagged genuine unintended behaviors, such as fiber optic interruptions in voltage meters and overtemperature shutdowns. Conversely, windows with low reconstruction errors were confirmed to be normal operations or standard low-wind shutdowns.

Conclusion

By combining 1D-CNNs with LSTM Autoencoders, this research demonstrates a robust and computationally efficient method for predictive maintenance in wind energy. The system effectively identifies complex, multidimensional outliers in noisy industrial IoT data, providing a practical early warning system for frequency converter failures. While the current study focuses on the converter, future work will explore integrating data from other turbine components, such as gearboxes and blades, for a more holistic monitoring solution across different geographical locations.

Summary updated

BibTeX
@inproceedings{schwenzfeier2022outlier,
  author = {Nils Schwenzfeier and Markus Heikamp and Ole Meyer and Andre Hönnscheidt and Michael Steffes and Volker Gruhn},
  title = {{Outlier Detection in Wind Turbine Frequency Converters Using Long-Term Sensor Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21533},
}
Figure