Research on Wind Turbine Fault Detection Based on the Fusion of ASL-CatBoost and TtRSA

Lingchao Kong, Hongtao Liang, Guozhu Liu, Shuo Liu2023 · DOI

TL;DR

This paper introduces a novel approach for detecting faults in wind turbines, specifically addressing the challenge of imbalanced datasets and the difficulty of hyperparameter tuning. By combining an improved CatBoost algorithm (ASL-CatBoost) with a new optimization strategy (TtRSA), the researchers achieve higher accuracy and efficiency in identifying wind turbine faults.

What Problem Does This Paper Address?

Wind turbines are crucial for sustainable energy, but they are prone to various operational faults. Traditional fault detection methods often struggle with imbalanced datasets (where normal operation data far outweighs fault data) and rely on manual, time-consuming hyperparameter tuning, which can lead to suboptimal performance. This paper tackles these issues by proposing a more robust and automated fault detection system.

Methods

The core of this research lies in two main innovations:

  • ASL-CatBoost Algorithm: This is an enhancement of the CatBoost algorithm designed to better handle imbalanced datasets. The key modification is the replacement of the standard cross-entropy loss function with an "Asymmetric Loss" (ASL) function. ASL decouples the weights for positive and negative samples and uses a probability transfer mechanism to reduce the impact of negative samples, allowing the model to focus more on learning from difficult-to-detect fault data.
  • TtRSA (Tent chaotic map and t-distribution-based Reptile Search Algorithm): This is a novel meta-heuristic optimization algorithm developed to automatically tune the hyperparameters of the ASL-CatBoost model. TtRSA improves upon the standard Reptile Search Algorithm (RSA) by incorporating:
  • Tent Chaotic Mapping: For more uniform initialization of the search population, reducing the risk of falling into local optima.
  • t-Distribution Mutation Strategy: To adaptively adjust the search range, promoting global exploration in early stages and local exploitation in later stages.

Datasets

The study utilizes SCADA system data from two wind turbines, F1 and F2, provided by Goldwind. This dataset includes various operational parameters and is labeled for normal operation, icing faults, and invalid states. The F1 dataset was used for training and validation, while the F2 dataset served as the test set.

Key Insights and Ideas

  • Asymmetric Loss for Imbalance: The ASL function is crucial for improving fault detection accuracy when dealing with imbalanced datasets, as it allows the model to learn more effectively from the minority fault class.
  • TtRSA for Efficient Optimization: The TtRSA algorithm significantly improves the efficiency and effectiveness of hyperparameter tuning compared to manual methods or other standard optimization algorithms. Its ability to balance global exploration and local exploitation is key.
  • Synergistic Improvement: The combination of ASL-CatBoost and TtRSA yields a powerful system that not only detects faults more accurately but also automates the complex process of optimizing the model's performance.

Key Results

The proposed TtRSA-ASL-CatBoost algorithm demonstrated superior performance:

  • Improved Accuracy: The ASL-CatBoost algorithm, when optimized by TtRSA, achieved a higher accuracy (0.950136) and F1-score (0.949581) compared to the original ASL-CatBoost (0.949427 and 0.946341, respectively) and other benchmark algorithms like GBDT, XGBoost, LightGBM, and LSTMAE.
  • Faster Training: The TtRSA-optimized ASL-CatBoost algorithm achieved its high accuracy in a significantly shorter training time (8 minutes 52 seconds) compared to other optimized algorithms.
  • Robustness: Experiments with varying numbers of features showed that the ASL-CatBoost model, when optimized, maintained good performance across different feature sets, indicating enhanced robustness.

Limitations and Future Work

The authors acknowledge the following limitations:

  • Binary Classification: The current study focuses on a binary classification problem (fault vs. normal).
  • Multi-Class Fault Identification: The model does not currently distinguish between different types of faults in a multi-classification scenario.

Future work suggested includes:

  • Developing the algorithm to accurately identify and categorize multiple fault states.
  • Further optimizing the algorithm to improve accuracy and efficiency in more complex scenarios.

Summary updated

BibTeX
@misc{kong2023research,
  author = {Lingchao Kong and Hongtao Liang and Guozhu Liu and Shuo Liu},
  title = {{Research on Wind Turbine Fault Detection Based on the Fusion of ASL-CatBoost and TtRSA}},
  year = {2023},
  doi = {10.3390/s23156741},
}
Figure