Data-driven aerodynamic instabilities detection in centrifugal compressors

Abstract

Centrifugal compressors are machines of utmost importance in numerous industrial and high-tech applications. They are known to be prone to the appearance of aerodynamic instabilities at low mass flow rates, when operating close to peak performance. Instabilities are a number of flow structures that negatively impact the compressor. Their effects range from efficiency loss for inlet recirculation, through increased level of vibrations and risk of fatigue damage for rotating stall up to an abrupt machine destruction for surge.

Quick and accurate instabilities detection is a challenge. Detection of surge is often a top priority as it has the biggest consequences for the machine operation, however detecting other instabilities is also important for overall performance and long-time operability. A promising approach to detection is based on datadriven techniques, using high frequency signals sampled from the compressor to capture the dynamics of the system. Such approach could warn about the approaching onset of instability, providing ample of time for reaction. However, the signal is often composed of a number of overlapping sources and a considerable amount of noise, which makes it a challenge to extract the meaningful indication of instability. A valuable insight into the system state could be obtained if the sources and the noise were separated.

The aim of this thesis is to build an instabilities-detection methodology leveraging data-driven signal decomposition techniques. The goal is to use a pressure signal collected inside of the compressor and obtain a real-time indication of the compressor stability. Two distinct decomposition methods, Empirical mode decomposition (EMD) and singular spectrum analysis (SSA) are investigated for this purpose. The goal of each of the method is to provide components sensitive to the presence of individual instabilities to build instabilities-sensitive features. The features are combined in the feature space, dimensionality of which can be adjusted depending on the system under analysis and expected unstable conditions. Using the decomposition techniques it is possible to increase the dimensionality of a signal, enabling differentiation of different types of instabilities present in the signal that would otherwise provide an overlapping signature in the original signal.

The proposed methodology is validated with the data from a low-pressure industrial compressor, equipped with five high-frequency pressure transducers located along the flow path. The compressor was operated through a wide spectrum of conditions. In the post-processing, the data was divided into different general conditions, being stable, locally unstable and globally unstable.

The results highlight the potential of defining robust features using both EMD and SSA for detecting general conditions, even with a relatively short input signal. The features are physically interpretable, and it is possible to provide meaningful thresholds for the detection of instabilities based solely on stable conditions. This is an important advantage, as operating the compressor in an unstable range brings risk of its damage. The overall accuracy of both methods is over 90%, with the majority of misclassifications coming from the region where the conditions transition from locally unstable to globally unstable. For certain machines, the extension of the operating range at the expense of safety might be beneficial. The globally unstable conditions reported in the case study can be furtherly divided into transient and deep surge. It is shown that decoupling those two instabilities for a robust indication with either EMD or SSA is not fully possible, which may come from the physical character of each instability. The features values for unstable conditions have to be known to differentiate transient and surge, hence the benefit of relying solely on stable data is lost. Obtaining features sensitive to each instability requires a longer input signal and extended processing, which negatively affects the responsiveness of the detection system. To avoid such issue, it is possible to use a general condition feature. It also requires prior mapping, but a robust indication can be obtained with a short input signal.

The values of features obtained from the process show certain level of variability and tend to overlap due to noise present in the data. With a prior mapping needed for the detection of exact instabilities, a probabilistic approach to classification can be leveraged. Apart from classification, such approach provides an information about the probability of a given class, which can be used to define no-classification zones in the feature space, where the probability of each of the classes is low. It is shown that the application of probabilistic model provides comparable classification rate, but it can offer increased flexibility and limit the number of sensors to be used for detection.

The approach demonstrated in this thesis can enable better understanding of the compressor operating conditions in the proximity of the surge line. Consequently, it could be useful for ensuring that the machine can safely reach its peak performance, possibly extending its operating range for different conditions.