Intelligent Steam Power Plant Boiler Waterwall Tube Leakage Detection via Machine Learning-Based Optimal Sensor Selection

Salman Khalid, Woo Cheol Lim, Heung Soo Kim, Yeongtak Oh, Byeng D. Youn, Hee-Soo Kim, Yong-Chae Bae2020 · DOI

TL;DR

This paper tackles the critical issue of waterwall tube leakage in steam power plants (SPPs) by proposing a novel machine learning approach that first optimizes sensor selection. By intelligently identifying the most informative sensors, the system achieves higher accuracy in detecting leaks while significantly reducing computational complexity.

What Problem Does This Paper Address?

Steam power plants are vital for electricity generation, but boiler failures, particularly waterwall tube leaks, are a major cause of outages and can significantly impact safety and economics. Early detection of these leaks is crucial for efficient maintenance and preventing costly downtime. Traditional methods for detecting these leaks often rely on complex models or expert knowledge, and multivariate algorithms, while powerful, can be hindered by the sheer volume of sensor data, including redundant and irrelevant information. This redundancy can degrade performance and increase computational burden. The paper addresses the challenge of efficiently and accurately detecting waterwall tube leaks by developing a machine learning model that first intelligently selects the most sensitive sensors, thereby overcoming the limitations of using all available data.

Methods

The proposed methodology is a four-stage process:

  1. Data Acquisition: Gathering relevant sensor data from the power plant.
  2. Data Preprocessing: This crucial step involves wavelet denoising to remove noise from the sensor signals. Wavelets are chosen for their ability to handle non-stationary signals and analyze data in both time and frequency domains, which is advantageous when noise sources are unknown.
  3. Optimal Sensor Selection: This is the core innovation. The paper employs Pearson correlation analysis to identify and remove redundant and irrelevant sensors. Sensors with a high correlation coefficient (greater than 0.95) are considered redundant, and only the most informative ones are retained. This reduces the dimensionality of the data while preserving critical information.
  4. Machine Learning Classification: Four supervised machine learning algorithms—Support Vector Machines (SVMs), k-Nearest Neighbors (k-NNs), Naïve Bayes (NB), and Linear Discriminant Analysis (LDA)—are trained and evaluated using both the raw sensor data and the optimally selected sensor data. Statistical time-domain features (RMS, variance, skewness, kurtosis) are extracted from the sensor signals to further enhance the models.

Datasets

The study utilizes data from a real-world steam power plant. Specifically, it employs data from 38 sensitive sensors, including generator active power, tube metal temperatures (from thermocouples on superheaters and reheaters), and steam temperatures. For training and evaluation, ten days of data from a healthy boiler and ten days of data from a leakage event were used, totaling 864,000 data points for each state. The dataset was split into 80% for training and 20% for testing.

Key Insights and Ideas

  • Sensor Redundancy is a Major Hurdle: The paper highlights that while many sensors are installed in power plants, a significant portion of their data can be redundant or irrelevant for specific tasks like leak detection. This redundancy can negatively impact the performance and efficiency of machine learning models.
  • Wavelet Denoising is Effective for Noisy Power Plant Data: The use of wavelet denoising proved effective in cleaning up noisy sensor signals, which are common in harsh industrial environments. This step is critical for accurate subsequent analysis.
  • Correlation Analysis is a Simple Yet Powerful Feature Selection Tool: Pearson correlation analysis emerged as a straightforward and computationally efficient method for identifying and removing redundant sensors. This approach is particularly suitable for industrial applications where computational cost and ease of implementation are important.
  • Optimal Sensor Selection Enhances ML Performance: By reducing the number of input sensors from 38 to 21 (a 44% reduction), the machine learning models not only became more computationally efficient but also showed improved accuracy in detecting waterwall tube leaks.

Key Results

The proposed SVM-based integrated optimal sensor selection process achieved the highest accuracy.

  • Raw Dataset (38 sensors): SVM achieved a testing accuracy of 88.2%.
  • Optimal Dataset (21 sensors): SVM achieved a testing accuracy of 90.5%.

This represents a significant improvement in accuracy (2.3% increase for SVM) after implementing the optimal sensor selection. Similar improvements were observed for k-NN, NB, and LDA classifiers. The confusion matrices also indicated that the models were better at correctly identifying healthy states but still showed improved performance in detecting leakage states with the optimized sensor set.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the correlation analysis effectively reduces redundancy, it might not always identify the absolute "best" set of sensors if complex, non-linear relationships exist between sensors that are not captured by pairwise correlation. The study focused on a specific type of boiler (CFB boiler), and its direct applicability to other boiler types might require further validation.
  • Future Work: The authors suggest that this simple and clear optimal sensor selection technique is quick and easy to implement in SPPs. Future work could explore more advanced feature selection techniques or investigate the application of this methodology to other types of faults or industrial processes.

Summary updated

BibTeX
@misc{khalid2020intelligent,
  author = {Salman Khalid and Woo Cheol Lim and Heung Soo Kim and Yeongtak Oh and Byeng D. Youn and Hee-Soo Kim and Yong-Chae Bae},
  title = {{Intelligent Steam Power Plant Boiler Waterwall Tube Leakage Detection via Machine Learning-Based Optimal Sensor Selection}},
  year = {2020},
  doi = {10.3390/s20216356},
}
Figure