Wind turbine failure prediction and health assessment based on adaptive maximum mean discrepancy

Jieyang Peng, Andreas Kimmig, Zhibin Niu, Jiahai Wang, Xiufeng Liu, Dongkun Wang, Jivka OvtcharovaInternational Journal of Electrical Power & Energy Systems 2021 · DOI

TL;DR

This paper introduces a novel framework for predicting wind turbine failures and assessing their health using an adaptive maximum mean discrepancy (AMMD) method combined with a convolutional neural network (CNN). This approach significantly improves failure prediction rates and reduces false alarms compared to traditional methods, offering a more interpretable and accurate system for predictive maintenance.

What Problem Does This Paper Address?

Wind turbines are complex machines with high operational and maintenance costs. Failures can be catastrophic, leading to significant downtime and expensive repairs. While condition monitoring and early warning systems are crucial for reducing these costs, existing data-driven approaches often struggle with high-precision failure predictions and interpretable health assessments. Specifically, traditional methods can generate a high number of false alarms, especially when dealing with the dynamic operating conditions of wind turbines and the sophisticated control systems that can make normal operations appear as anomalies. This paper aims to bridge this gap by developing a more robust and accurate system for predicting failures and assessing the overall health of wind turbines.

Methods

The authors propose a three-stage methodology:

  1. Data Acquisition and Pre-processing: SCADA data from wind turbines are collected, cleaned (outlier elimination), and normalized. Failures are then statistically categorized based on their frequency and downtime.
  2. Model Training:
    • Critical Failure Identification: A four-quadrant diagram is used to classify failures into categories, with a focus on "low frequency and extensive downtime" failures for predictive maintenance.
  3. Adaptive Maximum Mean Discrepancy (AMMD): This is the core of their novel approach. AMMD is an extension of the Maximum Mean Discrepancy (MMD) metric, which measures the similarity between probability distributions. In AMMD, weights are assigned to different sensor features, allowing the model to adaptively learn the importance of each sensor for fault prediction. These weights are optimized to maximize the prediction of critical failures.
  4. Convolutional Neural Network (CNN): A CNN is employed to identify and reduce false alarms generated by the AMMD. The CNN is trained to distinguish between true and false positive predictions.
  5. Failure Prediction: The trained AMMD-CNN model uses real-time data to assess the turbine's health and predict potential failures.

Datasets

The study utilizes a high-frequency SCADA dataset collected from six wind turbines in an offshore wind farm in eastern China. The data spans from January 24 to November 3, 2015, and includes 31 different sensor parameters. The dataset was divided into training, validation, and testing sets (70%, 10%, 20% respectively). No new benchmark datasets were introduced; rather, existing SCADA data was used for evaluation.

Key Insights and Ideas

  • Adaptive Sensor Weighting is Key: The AMMD approach, by adaptively assigning weights to sensors, allows the model to focus on the most relevant features for predicting specific failures, leading to improved accuracy.
  • CNN for False Alarm Reduction: The integration of a CNN is crucial for filtering out false positives, a common problem in traditional condition monitoring systems, thereby increasing the reliability of the alerts.
  • Categorizing Failures for Targeted Maintenance: Identifying critical failures (those with low frequency but high downtime) and tailoring maintenance strategies accordingly is essential for optimizing resource allocation and minimizing operational impact.
  • Interpretability through Feature Weights: The learned sensor weights in the AMMD model provide a degree of interpretability, indicating which sensors are most indicative of certain failure modes.

Key Results

  • Higher Failure Detection Rate: The AMMD-CNN model achieved a significantly higher failure detection rate compared to traditional methods like PC2-DEV, Bhattacharyya distance, and KL divergence. For example, recall rates for AMMD-CNN were consistently above 54%, reaching up to 81.2% for Turbine 9, while other methods struggled to exceed 28%.
  • Superior Precision: The precision of the AMMD-CNN model was remarkably high, often exceeding 99%, indicating a very low rate of false alarms. This is a substantial improvement over methods that reported precision rates as low as 2-8.5%.
  • Identification of Failure Types: The study successfully categorized failures into "identifiable and predictable," "identifiable but unpredictable," and "cannot be identified or predicted," providing insights into the limitations and strengths of the proposed model.

Limitations and Future Work

  • Data Limitations: The study acknowledges that the proximity of the wind turbines in the dataset means they share similar environmental conditions, service age, and performance. Future work should explore wind farms in diverse locations and with varying turbine characteristics (onshore/offshore, different ages, types) to generalize the findings.
  • Predictive Degradation Modeling: The authors suggest extending the research to infer failure progression from degradation model results, enabling proactive identification of failure attribution and calculation of remaining useful life.
  • Improving Prediction for Unpredictable Failures: For failures that were not successfully predicted, further investigation into which sensors could improve prediction success rates is recommended.
  • Stochastic Failures: The paper notes that stochastic failures, like those related to hydraulic oil depletion, can be challenging due to gradual degradation and the inherent delay in sliding window approaches. Further research into optimizing sensor weights or exploring alternative modeling techniques for such failures is warranted.

Summary updated

BibTeX
@article{peng2021wind,
  author = {Jieyang Peng and Andreas Kimmig and Zhibin Niu and Jiahai Wang and Xiufeng Liu and Dongkun Wang and Jivka Ovtcharova},
  title = {{Wind turbine failure prediction and health assessment based on adaptive maximum mean discrepancy}},
  journal = {International Journal of Electrical Power \& Energy Systems},
  year = {2021},
  doi = {10.1016/j.ijepes.2021.107391},
}
Figure