A Novel Intelligent Method for Fault Diagnosis of Steam Turbines Based on T-SNE and XGBoost

Zhiguo Liang, Lijun Zhang, Xizhe WangAlgorithms 2023 · DOI

TL;DR

This paper introduces a novel method for diagnosing faults in steam turbines by combining t-SNE for dimensionality reduction and K-means for initial fault identification, followed by SMOTE to handle imbalanced data and XGBoost for accurate multi-class classification. This approach achieves high accuracy (97%) and provides early fault warnings, crucial for preventing costly downtime in thermal power plants.

What Problem Does This Paper Address?

Steam turbines are critical components in thermal power generation, but their frequent failures can lead to significant financial losses. Early and accurate fault diagnosis is essential to minimize these risks. However, several challenges hinder effective diagnosis:

  • Data Imbalance: Real-world operational data often has a severe imbalance between normal and fault states, with fault instances being much rarer. This makes it difficult for machine learning models to learn effectively from limited fault data.
  • High-Dimensional Data: Sensor data from steam turbines can be high-dimensional, making it computationally expensive and challenging to analyze directly.
  • Data Labeling: Obtaining accurately labeled fault data can be difficult due to manual recording processes, missing information, or blurred records.
  • Non-linearity and Non-stationarity: The signals collected from turbines are often non-linear and non-stationary, making traditional signal processing methods insufficient for extracting meaningful fault features.

This paper aims to address these challenges by proposing an intelligent method that can effectively extract features, handle data imbalance, and accurately classify faults in steam turbines using real-world operational data.

Methods

The proposed method is a multi-stage process designed to overcome the aforementioned challenges:

  1. Dimensionality Reduction and Initial Clustering (t-SNE & K-Means):
    • t-SNE (t-distributed Stochastic Neighbor Embedding): This non-linear dimensionality reduction technique is used to map the high-dimensional sensor data into a lower-dimensional space (specifically, 2D in this study). t-SNE preserves local data structure, meaning similar data points in the high-dimensional space are likely to remain close in the low-dimensional representation. This makes the data more amenable to visualization and subsequent clustering.
  2. K-Means Clustering: Applied to the low-dimensional data from t-SNE, K-means is used to group data points into clusters. This step helps in distinguishing between normal and potentially faulty operational states, providing an initial label for the data.
  3. Handling Data Imbalance (SMOTE):
    • SMOTE (Synthetic Minority Over-sampling Technique): After initial labeling, the significant imbalance between normal and fault data is addressed using SMOTE. This algorithm generates synthetic minority class samples (fault data) by interpolating between existing minority class samples and their nearest neighbors. This balances the dataset without simply duplicating existing data, which can lead to overfitting.
  4. Multi-Class Fault Classification (XGBoost):
    • XGBoost (Extreme Gradient Boosting): This powerful ensemble learning algorithm is employed for the final multi-class classification task. XGBoost is known for its efficiency, scalability, and ability to handle complex datasets. It builds an ensemble of decision trees, iteratively improving predictions by focusing on misclassified instances. The paper highlights XGBoost's advantages in preventing overfitting through regularization and its computational efficiency.

Datasets

The study utilizes a real-world dataset from Steam Turbine 2 in a thermal power plant in China.

  • Data Source: The data comprises time-series data from the Supervisory Control and Data Acquisition (SCADA) system (sampling rate < 1 ms, interval sampling of 1 second) and fault information from manual records and the system application and product (SAP) system.
  • Data Characteristics: The dataset contains 34 variables, including timestamps, operating condition parameters, and status parameters. The acquisition period spanned eight months, resulting in approximately 340,000 data points.
  • Fault Data: Five types of faults were identified and used for analysis. The fault information was filtered from manual records to ensure it represented actual equipment operation faults.
  • Comparison to Existing Benchmarks: The paper emphasizes the use of actual plant data rather than experimental or simulated data, which is a key differentiator from many existing studies. The data length is also significantly larger (months to years) compared to typical experimental datasets (hours to days).

Key Insights and Ideas

  • Synergistic Feature Extraction: The combination of t-SNE and K-means effectively tackles the challenge of extracting meaningful features from high-dimensional, unlabeled operational data. t-SNE provides a visually interpretable low-dimensional representation, and K-means leverages this to perform initial fault detection.
  • SMOTE's Crucial Role: The paper underscores the critical importance of SMOTE in addressing the severe data imbalance inherent in real-world fault diagnosis. Without it, the performance of the subsequent classifier would be significantly hampered.
  • XGBoost's Robustness: XGBoost proves to be a highly effective classifier for this complex multi-class problem, demonstrating superior performance compared to SVM and Random Forest in their comparative analysis. Its ability to handle complex interactions and prevent overfitting is key.
  • Early Warning Capability: The method not only diagnoses faults but also provides an early warning, with the potential for at least two hours of advance notice. This is invaluable for proactive maintenance planning.
  • Practical Applicability: The research focuses on actual power plant data, making the proposed method directly applicable and valuable for industry, rather than being limited to theoretical or simulated scenarios.

Key Results

  • Overall Accuracy: The proposed method achieved an impressive overall accuracy of 97% on the test set.
  • Early Warning: The system demonstrated the capability to provide an early warning of at least two hours in advance of a fault.
  • Classification Performance: Precision, recall, and F1-score for individual fault classes were generally high, with values often exceeding 95%, indicating robust classification performance across different fault types. For instance, Fault 1 achieved a recall rate of 100%.
  • Computational Efficiency: While t-SNE can be computationally intensive, the training time for the entire model (including dimensionality reduction) for approximately 340,000 samples was less than one hour, which is considered acceptable for enterprise-level systems. Real-time prediction is achieved within 1 second.

Limitations and Future Work

  • Computational Efficiency of t-SNE: The paper acknowledges that t-SNE's dimensionality reduction can be computationally expensive for extremely large datasets, potentially taking several hours. However, they deem it acceptable for their current dataset size and enterprise application context.
  • Algorithm Improvement: While the current algorithm is effective, the authors suggest that for real-world applications, continuously updating the training model with more historical data (e.g., seven years of data) has shown good results.
  • Generalizability: The paper mentions validating the method on pneumatic feed pump data, suggesting its applicability to other equipment within a power plant. Future work could involve broader validation across diverse equipment types.
  • Data Uncertainty: The paper notes that the initial data might have missing or blurred fault records, which were filtered. Further research could explore methods to robustly handle such data imperfections.

Summary updated

BibTeX
@article{liang2023novel,
  author = {Zhiguo Liang and Lijun Zhang and Xizhe Wang},
  title = {{A Novel Intelligent Method for Fault Diagnosis of Steam Turbines Based on T-SNE and XGBoost}},
  journal = {Algorithms},
  year = {2023},
  doi = {10.3390/a16020098},
}
Figure