TL;DR
This paper introduces a novel Support Vector Regression Machine (SVRM) model to predict fan rotation stalls in power plants using in-situ pressure sensor data. By accurately predicting internal fan pressures, the model enables early detection of stalls, allowing for timely intervention to prevent costly failures and improve energy efficiency.
What Problem Does This Paper Address?
Thermal power plants rely heavily on large blowers and exhaust fans, which consume a significant portion of their electricity (over 30%). Failures in these fans, particularly due to rotation stalls, are a leading cause of unscheduled shutdowns, resulting in substantial economic losses. Rotation stalls occur when airflow within the fan impeller is disrupted, leading to increased loads, potential blade fatigue, reduced efficiency, and even catastrophic damage. Existing methods often detect stalls only after they have begun, limiting the ability to actively prevent them. This paper aims to bridge this gap by developing a predictive model that can anticipate stalls before they occur, enabling proactive control measures.
Methods
The core of this research is a Support Vector Regression Machine (SVRM) model. To prepare the data for the SVRM, the authors employed a chaotic phase space transformation technique. This involves reconstructing the system's dynamics from time-series pressure data into a higher-dimensional space, which is believed to capture more comprehensive information about the system's behavior.
For predicting rotation stalls, the SVRM model is trained on experimental pressure data. Once trained, it predicts future pressure values. These predicted pressure signals are then analyzed using wavelet transform, a signal processing technique adept at identifying frequency components over time, especially in non-stationary signals like those found during stall development. Specifically, the Daubechies orthogonal wavelet (db4) was used for multi-resolution decomposition. The fourth-layer detail coefficient, which captures frequencies related to the rotating stall, was used as a quantitative indicator. A threshold was set on this coefficient to detect the onset of a stall.
Datasets
The study utilized experimental data collected from a 4-73No8D centrifugal fan, a common model in Chinese thermal power plants. Pressure signals were measured under various operating conditions, including normal operation and transitioning into rotation stall. The data was collected at a sampling frequency of 320 Hz and later reduced to 80 Hz to investigate multi-step prediction capabilities. The experiments involved varying guide vane openings (0°, 15°, 30°, and 45°) and controlling airflow resistance with a throttling cone to induce stall conditions.
Key Insights and Ideas
- Predictive Power of SVRM: The SVRM model, when combined with phase space reconstruction, demonstrates a strong capability to predict future pressure behavior, moving beyond simple detection to actual prediction.
- Wavelet Transform as a Stall Indicator: The fourth-layer detail coefficient from the Daubechies wavelet transform effectively isolates and quantifies the energy associated with the rotating stall frequency, making it a reliable indicator.
- Early Warning System Potential: The ability to predict stalls with a time advance of up to 0.0625 seconds is crucial. This lead time is significant enough for control systems to react and prevent stall development, a key advantage over existing methods.
- Centrifugal vs. Axial Fans: The paper highlights that research on rotation stalls in centrifugal fans is less extensive than for axial flow compressors, and there are significant differences in their stall mechanisms, underscoring the importance of this work.
Key Results
The SVRM model successfully predicted fan pressure data with a time advance of up to 0.0625 seconds. This prediction capability was validated by comparing stall detection results from both measured and predicted pressure data using wavelet transform. The predicted stall onset points closely matched the actual stall onset points derived from measured data, demonstrating the model's reliability. Single-step predictions were highly accurate, and even multi-step predictions (up to five steps) yielded reliable results for stall prediction. Six-step predictions, however, showed accumulated errors that compromised accuracy. The average computation time for the SVRM model was a mere 0.00096 seconds per prediction.
Limitations and Future Work
The paper acknowledges that multi-step prediction errors accumulate, which limits the number of reliable prediction steps. While five steps proved effective, six steps led to inaccuracies. The authors suggest that for online stall prediction, time-domain stall detection techniques could be integrated with the SVRM's predicted data. Future work will focus on developing active control mechanisms that utilize this predictive capability to actively suppress fan rotation stalls, further enhancing the safety and efficiency of power plant operations. The model's potential application to other types of fans and in combustion-based power plants is also highlighted.
Summary updated