A Review of Stall Detection in Subsonic Axial Compressors

Abstract

Stall events in axial compressor systems have been a limiting factor for efficiency of such systems and a source of safety concerns. The detection of the onset of stall, and in many cases the precursor of the onset of stall, have been of interest in the axial compressor community for many decades. As such, development of algorithms along with active control could lower cost, reduce emissions, improve safety, and increase market competitiveness. To gain an understanding of these stall phenomena, past and current research has focused on modeling axial compressors as dynamic systems, with a focus on obtaining descriptive formulations of the physical aspects of stall. Some of these approaches allow for active control measures that extend the stall margin of the compressor system to increase safety and efficacy. This paper reviews the major contributions in these listed pursuits and presents the latest methods and algorithms for stall precursor detection in low-speed axial compressors. In particular, a review is presented in the types and characteristics of stalls, the major mathematical models used to describe these systems, influences of physical attributes such as tip clearance, guide vanes, and groove casing—operating as passive control elements—but also active control utilities such as air injection are discussed along with a detailed review of existing stall precursor detection algorithms. In addition, a forward-looking projection is presented that includes the use of machine learning algorithms to further the understanding and the capability of stall precursor detection.

1. Introduction

This review focuses on the efforts to characterize flow and flow phenomena in axial compressor systems. As safety is a predominant criterion in the engineering of these types of systems, characterization of its internal dynamics and their associated impact on safety have been of great interest in the research as well as the commercial community. Safety infractions are partially observed by the onset of stall. These types of stalls are accompanied by a rather sudden loss of pressure, and is often at an engine’s operating point that is near its optimum in terms of efficiency. Hence, the design point for such aeroengines is below its optimum, giving rise to the term stall margin. Once an engine experiences stall, to un-stall the engine, the flow rate is reduced considerably and recovery is associated with a hysteresis loop, which consumes more time than usually is available when operated in flight. The ensuing stall rotates with the compressor rotation at a lower rate and affects a section of blades, dividing the compressor circumference into a stalled section and an unstalled section. Stall inception is speculated to originate by long-scale disturbances—or modal stall or by short-scale disturbances—or spike stall. An accepted roadmap to stall was proposed in 2009 by Lin et al. [1], where the momentum ratio—as measured as the linear momentum of the tip leakage flow across the blade gap to the linear momentum of the main flow—versus the relative tip gap maps the different states of flow from stable and steady flow to stable yet unsteady flow, to eventually unstable flow. Research into the disturbance phenomena that compressors face includes blade flutter and vibration. Research has shown in Ref. [2] that blade flutter is affected by blade damping and can affect stall. Shown in Ref. [3], flutter can cause stall incident, specifically in supersonic, but may be seen in subsonic cases.

Stall and surge are instabilities that compressors inherently possess. The possibility of running compressors into either stall or surge causes major safety concerns. When fluid is flowing through the compressor, the flow should be smooth. The first-time observations of stall were made when water was used as the fluid, allowing disturbances to be more visible. Over the years, research has been focused on both stall and surge, as they are a limiting issue when it comes to compressor design [4].

Compressor systems are categorized in terms of compressor maps. These maps consider the physical size, the number of stages, the inclusion of inlet guide vanes, and other physical aspects. Each compressor has a unique map that influences the behavior of the system. Experimentally collected data are used to create these maps that involve the mass flow rate and the pressure rise or change in pressure. Each compressor, even among identical models, can have variations in the compressor map resulting from manufacturing tolerances. The maps are essential due to their ability to indicate the location of stall, but their creation is difficult to do by experimentation alone, as a compressor cannot safely be ran into stall during data collection. This means there is a need to both interpolate and extrapolate data [5].

A sample compressor map is shown in Figure 1. The x-axis is mass flow rate and the y-axis is change in pressure or pressure ratio. The map has lines of constant speed and efficiency. Figure 1 shows an average operating line, which lies a distance from the stall line called the stall margin. The stall margin allows for safe operation to occur with a factor of safety away from the stall line.

Research has been done on compressor maps, as they are a vital part of describing compressor systems. Gholamrezaei and Ghorbanian [6] investigated how to use neural networks to better describe the maps, especially for maps built from limited experimental data.

Compressor map representations commonly have the x- and y-axis normalized. Moore and Greitzer [7] developed a set of nonlinear third-order partial differential equations where Φ is the compressor mass flow, Ψ is the pressure difference, and U is mean rotor velocity. This set of normalized equations results in a cubic representation of the compressor curve, as shown in Figure 2. This figure shows how the negative slope sections are stable, solid green line in the green shaded area, while the positive slope is negative, dashed red line in the red shaded area. It also shows how an operating point is held away from the point where the system becomes unstable. This space is called the stall margin. The hysteresis is also shown in dark blue line.

Models that are capable of showing stall and surge are a major focus of research. Simulations created using the aspects of a physical compressor system are advantageous in research due to their ability to model when stall occurs without running a test compressor into destruction. The driving force of this approach comes from Moore and Greitzer [7]. They created a mathematical model that captures the dynamics of a compressor using physical geometry. This model is capable of showing rotating stall and surge, which was a significant step forward in the ability to analyze the dynamics within compressors [7].

The efficiency of compressors increases the closer to stall it operates. Due to stall being an often destructive instability, it is important to safely sit below stall with the operation point. This buffer reduces the risk of compressors entering stall. Controllers are designed to keep the operating point away from the stall line while working towards the highest efficiency possible. It is favorable to use simulation in both the design process and testing.

Figure 3 shows rotor and stator of stage 10 and rotor of stage 11 of a J34 Westinghouse jet engine with the corresponding stator. The J34 Westinghouse jet engine has an 11-stage compressor section.

Single-stage axial compressors have a low pressure rise due to the small rate of change of the cross-sectional area. For most practical uses, a multistage compressor is used. In this case, the cross-sectional area decreases through the stages. The rotor transfers power to the fluid and the stator transforms that kinetic energy into an increase in static pressure [8].

Instability occurs when the system is throttled to the stall limit. Two phenomena are produced, called rotating stall and surge. Rotating stall involves flow separating from the blade. As the flow separates from the surface of the blade, disturbances from that cause a decrease in flow through the blade passage. Air is diverted from the original passage into adjacent passages. With this increase in air flow, disturbances are caused in those passages that disrupt the flow and cause blockages. In this way, the stall cell propagates along the row of blade passages [8].

Stall can occur with at least two distinct types that include full-span and part-span. The difference between the two is how much of the annulus is stalled. In full-span, the whole annulus is stalled, while in part-span, only a portion of the blade passage experiences rotating stall. Full-span is more common in axial compressors that have a high hub/tip ratio. Part-span is more common in high-speed multistage compressors [8].

Stall causes numerous issues. The first being that the pressure rise is greatly reduced when there are stalled cells. This causes an efficiency issue, as the compressor is not performing as needed. The more alarming issue is that it can cause vibration in the cell passages due to the stalled cells moving at a slower rate than the non-stalled cells. Mechanical failure can occur, especially if the vibration frequency matches the natural frequency of the blade. To clear stall, the throttle can be used. However, the throttle has to be opened past its previous position [8].

Surge is another instability that compressors have. Surge is characterized by flow oscillation through the compressor that may limit the compressor characteristics. This oscillation can cause a loss in efficiency as well as damage to the compressor. The damage also occurs from vibration caused from surge, which affects components such as connected piping [8].

This review starts with a discussion about modal and spike stall behavior. It then covers modeling of compressors and how it relates to stall and surge occurrence. A review of the physical aspects of stall is given that includes tip clearance and grooved casing. This leads to a review on stall precursors, which are moments in the flow of a compressor that are indicative of impending stall. Finally, this review finishes with current and upcoming methods to find stall precursors. Stall precursors are important for the design of compressors, as it allows for active control of stall to increase efficiency and safety. This review strives to be a comprehensive review of aspects that affect the inception and detection of stall and stall precursors.

2. Modal and Spike Stall

Camp and Day [9] discuss spike and modal stall in their seminal 1997 award-winning paper. They explain the physical differences between the two phenomena, then with data gathered from an experiment, propose a model to explain the type of stall inception pattern observed in a particular compressor. They concluded that spike stall and modal stall are, respectively, caused by spike disturbances or modal oscillations. Spike stall earned its moniker due to the spikelike appearance in the air flow velocity traces. Spike stall is a 3D breakdown of the flow field due to high rotor incidence angles. When the stall cell appears, it travels quickly around the annulus at roughly 60% to 80% of the rotor speed. But as the stall cell grows, the rotation speed decreases. The short length scale disturbances and high initial speed of rotation are characteristic of a spike stall.

In Day’s 1993 paper [10], it is reported that modal perturbations/oscillations—sometimes referred to as ‘pre-stall waves’—are small sinusoidal velocity fluctuations that rotate about the annulus. These modal oscillations would start with a small amplitude that is indistinguishable from the background noise. As the instability point of the compressor is reached, the wave would grow rapidly and cause the compressor to stall. However, Day also noticed that if finite stall cells develop before modal oscillations, the symmetry of the flow field is destroyed, and modal oscillations do not manifest.

Returning to Camp and Day’s 1997 paper [9], modal oscillations are a 2D instability of the whole compressor system and have a long length scale disturbance. Modal oscillations can either develop smoothly into rotating stall when the low velocity trough initiates a flow breakdown over a wide section of the annulus, or a flow separation can appear near the tip of a blade row that produces a spike disturbance. Hence, modal oscillations can initiate both modal stall and spike stall. Modal oscillations typically have an amplitude of 2–3% of the free stream velocity when flow breakdown occurs. As also discussed by Day (1993) [10], modal oscillations are not necessarily a pre-stall cell. The oscillations appear close to the peak pressure rise characteristic, which has the rotor blades close to their stalling limits. Any oscillations at that point can then cause flow separation.

Camp and Day’s 1997 [9] experiment suggested that if the critical rotor incidence is to the left of the peak pressure characteristic, the compressor will experience modal stall inception. If the critical rotor incidence is to the right, then the compressor will experience spike stall inception.

Simpson et al., 2007 [11], determined that the onset of stall can be influenced by both the inlet flow and downstream stator characteristics. The ‘Zero Slope Criteria’, where modal stall onset is expected near the peak total-to-static pressure rise, and the ‘Critical Incidence Hypothesis’, which predicts spike stall when flow angle at the rotor tip exceeds a critical threshold, were analyzed. They determined that the critical incidence depends on the flow environment in addition to the rotor geometry. Therefore, the meridional acceleration coefficient is proposed as an influence on critical incidence. The issue of burst ingestion was discussed, and more investigation was requested to better understand the effects on the stall inception mechanism. They also concluded that modal stall inception can be delayed until after the peak pressure rise of the total-to-static pressure rise characteristic.

The understanding of compressor behavior, particularly regarding stall and surge phenomena, has evolved through increasingly sophisticated models. Early models primarily addressed axial flow instability under static conditions, but modern developments have integrated dynamic elements such as rotational speed variations and flow-induced instabilities. This review outlines key advancements in compressor modeling over recent decades, emphasizing their contributions to the detection and control of stall.

One of the seminal works in compressor modeling is the Moore–Greitzer (MG) model, which laid the foundation for analyzing surge and rotating stall in axial compressors. Initially, this model assumed constant compressor speed, focusing solely on flow dynamics and pressure oscillations. However, their work, “A Theory of Post-Stall Transients in Axial Compression Systems” [7], provides a theoretical framework for understanding the behavior of axial compressors during post-stall conditions. The paper details their derivation of the nonlinear equations to capture the interactions between pressure rise, flow coefficient variations, and rotating stall cells during transients, using a Galerkin approach to simplify angular dependencies. This allowed the model to represent the growth and decay of stall cells over time, emphasizing the complex coupling between rotating stall and surge under dynamic conditions.

Moore and Greitzer’s analysis [7] highlighted that rotating stall is generally quasi-steady and non-axisymmetric, while surge remains unsteady and axisymmetric. Their work provided critical insights into how these phenomena can evolve and interact during post-stall transients, offering guidance for designing more resilient compressor systems that can maintain stability beyond initial stall events. These findings laid the groundwork for further studies that expanded the MG model’s applicability to real-world scenarios, such as those by Gravdahl and Egeland, who introduced spool dynamics into the framework.

Jan Tommy Gravdahl and Olav Egeland extended this model in their work “A Moore-Greitzer Axial Compressor Model with Spool Dynamics” [12]. Their research introduced a critical modification: the addition of spool dynamics, represented by a new variable, the B-parameter, to account for time-varying angular speed. This innovation transformed the static three-state model into a dynamic four-state system, making it more applicable to real-world compressor operations, especially under varying speed conditions.

Through simulations, Gravdahl and Egeland [12] demonstrated how this extended model better captured the transitions between rotating stall and surge, especially under variable speeds. Their findings highlighted that lower B-values typically led to rotating stall, while higher values induced surge. This extension allowed for more accurate control strategies and a deeper understanding of the dynamic coupling between compressor speed and instability modes, particularly important for modern controllers dealing with variable-speed compressors.

Building on their axial compressor work, Gravdahl and Egeland [13] further explored surge control in centrifugal compressors. In their paper “Speed and Surge Control for a Low-Order Centrifugal Compressor Model”, they developed a low-order model incorporating both mass flow and spool rotational speed to control both speed and surge conditions. They introduced a close-coupled valve (CCV) to alter pressure drop, thereby stabilizing regions that are typically prone to instability. The integration of a Proportional Integral (PI) controller allowed for simultaneous control of rotational speed and surge suppression.

The novelty of this approach was in its combination of Lyapunov-based stability analysis with speed control mechanisms, ensuring that even in areas previously unstable due to surge, the compressor could operate stably. This work marked an important milestone in the control of surge phenomena in variable-speed compressors, offering a robust approach to mitigating surge while maintaining desired performance [13].

Ali Ghaffari et al. [14] contributed significantly to the simulation of surge and stall dynamics in axial compressors with their work “Axial Compressor Surge and Stall Simulation and Sensitivity Analysis”. Their study used advanced simulation techniques to examine the behavior of compressors under both constant and variable speed conditions. Ghaffari’s team focused on aerodynamic instabilities, noting that surge and rotating stall were directly influenced by load line positioning and speed.

A key contribution of their work was the introduction of an adaptive neuro-fuzzy inference system (ANFIS) to predict compressor performance across different speeds, utilizing limited experimental data to build comprehensive performance maps. Additionally, the paper highlighted the importance of plenum volume in dictating system behavior, showing that larger plenum volumes increased instability, particularly during transient operations like shutdown. These findings further underscored the complex interplay between flow dynamics, load conditions, and geometric factors in the onset of stall [14].

A more recent development in stall detection was presented by Chun-Ming Chen et al. in their paper “Visualization and Analysis of Rotating Stall for Transonic Jet Engine Simulation” [15]. Their work approached the issue of rotating stall in jet engines from a data-driven perspective, leveraging spatiotemporal visualization techniques to detect early signs of stall. The system they developed provided interactive, comparative visualizations of simulation data, enabling more efficient detection of rotating stall precursors by tracking vortex dynamics.

This framework allowed for the exploration of flow anomalies over time and across blade passages, which was critical in designing more effective stall detection algorithms. Chen’s work represented a shift toward using computational tools and visual analytics for real-time stall analysis, reducing the complexity and time required for detailed stall detection in large-scale engines [15].

Yamada et al., in their paper “A Study on Unsteady Flow Phenomena at Near-Stall in a Multi-Stage Axial Flow Compressor” [16], used high-fidelity simulations on Japan’s K computer to explore unsteady flow phenomena in multistage axial flow compressors. By focusing on the critical role of hub-corner separation in stall initiation, they demonstrated how vortex structures, particularly at the stator blades, could lead to rotating stall.

Their simulations confirmed that stall propagation was closely tied to complex vortex formations, which could spread from the leading edge of the rotor. These insights offered a more detailed understanding of the mechanics behind stall inception and the role of flow separation in multistage systems. This work has significant implications for the design of industrial compressors, helping to improve stall margins and overall stability [16].

The historical development of compressor modeling has led to increasingly sophisticated tools for understanding and controlling surge and stall phenomena. Early models like the Moore–Greitzer framework provided the foundation for modern control strategies, while extensions incorporating spool dynamics and surge control laws have enabled more precise control over compressor performance. Recent advances in computational simulation and visualization tools have further enhanced the ability to detect and mitigate stall, offering powerful tools for both research and industrial applications. The ongoing development of these models will undoubtedly continue to push the boundaries of what is possible in stall detection and compressor stability.

Modeling has proven to be a reliable way to analyze stall conditions in compressor systems. Numerical models can show when stall and surge will occur and what affects their behavior. In particular, models have shown that load line positioning, speed, plenum volume, vortex formation, and flow separation affects stall. With these types of models, design and control criteria can be established. To look more into the dynamics of the flow characteristics, it is important to look at tip clearance and leakage.

3. Tip Clearance and Tip Leakage

Physical flow characteristics are shown to affect stall. One of the main characteristics is tip clearance and flow leakage from the tip. Many studies have been done to analyze exactly how tip clearance affects stall and what adjustments can be made to alleviate the disturbances that can occur at the tip. The following studies have been collected to discuss this issue in detail.

One of the key findings from investigating the flow is that the stall of modern fan rotors often emerged from the development of instabilities in the tip region. This is dependent on the strong interaction between the tip leakage vortex and the in-passage shock system. The importance of tip clearance on the stability of axial flow is that increasing the clearance between the tip of rotors and the casing of a low-speed compressor caused the point of stall inception to move to a higher flow rate. The tip clearance also plays a major role in setting the stall limit.

Adamczyk et al. [17] utilized the numerical simulation code, including the artificial damping operator and the turbulence model from Adamczyk et al. (1989) [18], to understand the effect of tip clearance on the near-stall end-wall flow field in a high-speed rotor by computing pressure and efficiency maps for four different clearance to tip chord ratios of 0, 0.25, 0.75, and 1.25 for Rotor 67. They observed that increasing tip clearance led to increased onset of stall, reduced flow range, and increasing stability. Zero tip clearance was better due to the absence of leakage vortex, and non-zero tip clearance exhibited instability due to leakage vortex and flow blockage.

Bae et al. [19] experimented with three different fluidic actuators to control the compressor tip clearance flow in a linear cascade using a low-speed wind tunnel. Four distinct criteria were found to influence blockage and loss linked to tip clearance, which include the type of actuator, its pitchwise location on the casing in relation to the location of the blade tip, its amplitude of actuation, and the frequency at which it is applied. The effectiveness of each active control approach was evaluated based on its ability to provide a reduction in tip leakage flow rate, an improvement in mixing between tip leakage and core, and an increase in streamwise momentum of the flow in the end-wall region. For the active control, three types of actuators are utilized. They are a jet that is normal to the mean flow that does not add to the net mass flux, called a Normal Synthetic Jet (NSJ), a jet that is aligned with the mean flow called a Directed Synthetic Jet (DSJ), and a Steady Directed Jet (SDJ). The NSJ offers both mixing enhancement and leakage flow reduction, while actuators with the SDJ and DSJ offer streamwise momentum boost. The work implied that the streamwise momentum injection is much more effective in tip clearance flow control.

Leitner et al. [20] demonstrated the relationship between the tip flow leakage flow with the incoming flow in a linear compressor cascade experimentally by methodically varying chord tip clearance between 2% and 4%, changing the angle of incidence and Reynolds number. The movement of inked fluid elements in space and time is taken into consideration throughout the research, which is conducted utilizing a water channel. They observed that increasing the tip clearance made the tip leakage roll up more tightly into a tip cortex and the rolling position shift downstream. The presence of tip clearance in the incoming flow causes a boundary layer detachment, which interacts with the rolling tip leakage vortex and moves upwards when reducing the tip clearance width.

Reeder [21] compiled a literature review on “Tip Clearance Problems in Axial Compressors” in which he states that scraping of the boundary layer by rotating blades leads to overestimating the reduction of efficiency due to tip clearance near stall. Jefferson and Turner [22] mentioned that tip clearance of about 1% of the blade height tends to inhibit local stalling at the tip, and reduced clearance leads to aggravated stall. They performed shrouded blading tests by increasing the shrouding clearance by around one percent of the blade height and noted a steady decline in performance. They found that if a compressor blade is working under near-stalling conditions, then shrouding tends to aggravate the tendency to stall.

William [23] deduced that when the flow is below the design flow rate, the axial velocity distribution at the tip drops off rapidly and is attributed to tip stall, and when performing experiments on a single-stage axial compressor, found that near stall, there is an increase in tip loading with increased tip clearance.

Vo [24] investigated the axial compressor stability using single- and multiple-blade passage computations. At the lowest flow coefficient, there is a development of equilibrium solution in single-rotor blade passage that causes the growth of tip clearance flow blockage until it reaches the hub. This onset causes spike disturbances in multiple-blade passages. The single-rotor passage’s lowest flow coefficient is determined by zero mass flow across the pitch at the trailing-edge blade tip and the commencement of leading-edge tip clearance flow spilling underneath the blade tip. The two conditions can occur at various flow coefficients, but both are required for spike disturbances to arise.

Nie [25] examined the dynamic behavior of stall under the influence of inlet distortion, radial loading distribution, tip clearance flow, stage matching, and downstream oscillation from combustion chambers to identify the most sensitive flow instability region to actuate and improve the operating stability margin. In this experiment, radial loading distributions along the blade height are modified by adjusting the inlet guide vane (IGV) configurations, and the changes in stall inception that occur are observed. Steady properties, dynamic pressure traces before rotating stall generation, and angle of attack features over the blade height are all analyzed using three independent measuring techniques. The following were observed: modal Stall Inception, which are prone to global disturbances, are indicated by negative angles of attack around the tip regions prior to the rotor’s leading edge. Spike Inception is caused by positive angles of α1 around the tip regions prior to the leading edge of the rotor, leaving the compressor susceptible to local disturbances, which in turn causes spike stall inception.

Zhang et al. [26] studied the mechanism behind deterioration of stability in axial compressors due to distorted flows. It was found that the tip leakage vortex is aerodynamically overloaded in distorted sectors, but remains in clean sectors. As a result, they advance in the direction of the leading edge, become erratic, and spill out of the blade passage, creating disruptions that resemble spikes. The behavior of blades in the distorted sectors is different from the one in clean sectors. The disruption that enters the clean sector, originating due to spinning of the vortex off the leading edge, can be minimized. Compressor stability is affected when the distorted sector’s rotational speed coincides with the propagation speed of the disturbances, akin to resonance. Casing treatments and tip injection are two ways to improve stability in the axial flow compressors that operate in distorted incoming flows.

Wu et al., 2012 [27], confirmed their simulations by finding a tip secondary vortex in tip flow fields that was the main source of the nearly periodic variation of efficiency. It was also determined that mass flow rate was negatively correlated with the simulated active period of the tip secondary vortex (TSV) until the last stable point. They identified a characteristic hump from the casing pressure measurements and determined that the cause was the movement of the TSV. They determined that the formation of a TSV comes from the low-energy leakage fluid coming from the breakdown of tip leakage vortex (TLV) in adjacent passages. Once formed, the TSV movement causes a periodic variation in the near tip loading, which alters the strength of the TLV and TSV, creating a self-sustained unsteady flow oscillation in the tip flow fields.

Wu et al., 2014 [28], performed a Fast Fourier Transform and a Short-Time Fourier Transform analysis to better understand and evaluate experimental observations concerning the flow behavior near the casing during stable and stall onset conditions. Some of the key findings from their study include the identification of a characteristic hump in the pressure spectra below the blade passing frequency, indicative of rotating instability in the tip flow field. The numerical analyses were able to replicate the experimental results and indicate the formation of tip secondary vortex (TSV) as a critical mechanism for rotating instability.

Berdanier et al., 2017 [29], reviewed a three-stage axial compressor for three tip clearances within an operating envelope from 52 to 100% corrected speed. The limiting stage was identified as the first stage, and by viewing the total-to-static characteristics for that limiting stage, it was sufficient to determine the slope of the characteristic for the whole system as the stall point. Further research is required to determine if this observation is like other multistage compressors.

Wang and Wu (2020) [30] studied the effects of large tip clearance in a compressor. They determined that rotational instabilities due to the large tip clearance can persist across a wide range of conditions. They were able to detect the instabilities with broadband hump with side-by-side peaks below blade passing frequency for rotating instabilities and low frequency peaks for rotating stall. The shifting and oscillating behavior of the tip leading vortices were determined to be the effect causing the disturbance in the pressure data. They also determined that, close to stall, the forward airflow spillage at the blade’s leading edge produces small repeating vortices that add to the existing rotating instability that stay as localized disturbances rather than developing into stall cells. Under rotating stall conditions, the Leading-Edge Vortex pattern is disrupted and creates scattered vortices that increase reversed flow that then leads to the eventual forming of localized stall cells.

Tip clearance and tip leakage have been shown to be key factors in stall. As tip clearance increased, the onset of stall also increased. However, too small of a clearance where the boundary layer is scraped can reduce efficiency. One of the causes of this tip clearance instability is tip secondary vortex (TSV) caused by tip leakage vortex breakdown. There are conflicting reports of the progression of stall. Many reports suggest that spike stall starts as short-scale and goes to long-scale with increasing tip clearance. However, another study suggests that the opposite is true, in which increased tip clearance goes from modal to spike. As tip clearance is a physical aspect of compressors that affect stall, so is casing treatment. Casing treatment will be discussed in the following section.

4. Physical Aspects That Affect Stall

Gaining a better understanding of the tip leakage flow in axial compressors and its implication in stall events provided motivation to a number of researchers to conduct extensive studies starting in the 1960s. Some of the earliest works were done at NASA, such as reported in Refs. [31,32], who conducted systematic studies of how to inject or remove air from the casing in order to determine the resulting effects on the compressor’s performance, in particular, the stall margin and efficiency. Koch and Smith, in Ref. [33], extended this work to include studies involving casing inserts. This work laid the foundation for understanding how manipulating the airflow near the casing could influence compressor behavior. These investigations included active control measures, such as air injection, and passive control measures, where the geometry of the casing wall is utilized to manipulate the tip leakage flow. This section is with regard to the latter approach: utilizing geometry in order to induce a passive control mechanism that aids in improving the stall margin.

Bailey and Voit, in Ref. [34], focused on the effects of porous casings in an axial-flow compressor rotor’s operating range. The study used a single-stage compressor and examined how different porosities and configurations of the casing affected performance. They found that porous casings could significantly extend the stable operating range of the compressor, particularly at lower flow rates. Their findings were crucial in demonstrating the potential of passive casing treatments. NASA’s interest in this topic is well documented, as these studies led to further investigations, such as the one by Osborn, Lewis, and Heidelberg in 1971 [35]. This NASA investigation examined several porous casing treatments and their effects on stall limit and overall performance. The researchers tested numerous configurations, including different hole sizes, angles, and distributions in the casing. They provided quantitative data on how these treatments affected the stall margin, pressure ratio, and efficiency. The study by Osborn et al. in 1971 helped establish early design guidelines for casing treatments.

Significant progress in understanding the mechanisms behind the casing treatment effectiveness was achieved through the work by Prince, Wisler, and Hilvers in 1974 [36]. The researchers conducted detailed flow measurements and created visualizations to understand how casing treatments affected the flow field within the compressor. They identified key phenomena such as the recirculation of flow through the treatment and its impact on tip leakage vortices. This work provided vital insights into why casing treatments were effective in improving stall margin.

Generally, the casing treatment studies can be categorized by three phases, where the first phase focused on compressor stability, with the key finding that slot-type casing treatment significantly improves the stall margin compared to other configurations, such as axial-skewed and axial-radial slot configurations, circumferential groove configurations [37], and perforated and honeycomb casing in Ref. [35]. The most documented type of casing treatment is the groove casing [35,38,39,40,41,42,43,44,45]. However, any stall margin improvement through casing treatments comes with a reduced efficiency [37]. At that time, researchers understood the relationship between the blade loading and flow separation. This relationship was documented in Ref. [38], where they noticed that an increase in blade loading causes flow separation on the suction side close to the blade tip. They also presented new insight into the flow mechanism with casing treatment and the related momentum transformation.

From these early works, the realization of achieving an improved stall margin using casing treatment will result in a negative effect on the efficiency of the compressor. This knowledge was one of the motivations for the second phase, where researchers tried to gain a deeper understanding of the flow dynamics for different types of casing treatments. A key finding of this second phase in casing treatment research was the comprehension of the influence the casing treatment has on the blade loading and its related momentum transfer. In particular, the authors in Ref. [46] concluded that unsteady effects in the casing slots were of secondary importance. The mean flow from rear to front of the slots was the primary mechanism for improving the stall margin, and the casing treatment appeared to work by recirculating and energizing the low-momentum fluid that would otherwise cause blockage.

For the third phase, the current phase, the tip leakage flow has become a prime target of investigations, where casing treatments in relation to the tip leakage flow are investigated [47].

Specifically for low-speed axial compressors, there are four major types of casing treatments investigated:

• Circumferential grooves, which are the most commonly referenced casing treatments and appear in two types of configurations: (a) single circumferential grooves, and (b) multiple circumferential grooves. This type of casing treatment has been documented to improve the stall margin of the compressor by reducing near-casing blockage and altering the axial momentum balance [48].
• Skewed axial slots, which are grooves or channels that are cut into the compressor casing above the rotor blade tips. These slots are oriented at an angle (skewed) relative to the radial direction, rather than being purely axial or circumferential. By recirculating high-pressure air from the blade tip region back into the flow, these slots help to energize low-momentum fluid near the casing, which can enhance overall flow characteristics and reduce the likelihood of stall [49].
• Self-injection configuration are arrangements that involve channels or passages in the compressor casing which allow high-pressure air from downstream compressor stages to recirculate back to upstream regions with the objective of improving the compressor’s stability. The recirculated air primarily affects the tip leakage flow and tip separation vortex [50].
• Hollow structures: for the purpose of flow control, hollow structures are patterns or configurations incorporated into the compressor casing with the intent to affect the tip end-wall flow and improve the compressor performance by controlling the boundary layer of the flow [35].

The effectiveness of these casing treatments can vary depending on factors such as the axial location of the treatment, the depth of the treatment, and the number of treatments. In addition, the choice and design of casing treatments are critical for optimizing compressor performance. For example, studies have shown that the axial location of circumferential grooves can significantly impact stall margin improvement, with optimal locations typically found at 10% and 50% axial chord aft of the leading edge [47].

5. Stall Precursor

As has been discussed in the previous sections, stall detection is a vital component of compressor design. However, being able to detect stall as early as possible allows for control methods to be instigated. To achieve this, research has been conducted on different types of stall to determine stall precursors at the earliest possible moment. The following section reviews some of this research describing stall precursors.

There are various authors in the past who conducted study in predicting stall based on the decreasing periodicity of the parameter chosen for the stall prediction. Inoue et al. [51], in 1991, investigated the statistical characteristics of pressure fluctuations on the casing wall to find a precursor to rotating stall. Near the inception of rotating stall, highly fluctuating pressure on the casing wall is observed near the leading edge. Cross-correlation of the pressure fluctuation between a reference measuring point and any other point was obtained. The possibility of predicting stall was observed. Tahara [52], in 2007, proposed a stall warning index by performing cross-correlation. The correlation was observed to be degraded near the stall inception. They stated that there was possibility of predicting stall this way; however, the application to predict the stall is not shown.

In 1995, Tryfonidis [53] examined nine compressors. He used Traveling Wave Energy as stall warning. The compressors were consistently indicating a warning time of 100–200 rotor revolutions prior to the modal stall event. However, the proposed method was not applied for spike stall. In 1999, Day et al. conducted a study to explain the role of stage matching in determining which stall would occur in a particular situation [54]. They performed data analysis and used spatial Fourier decomposition of modes and traveling wave energy analysis as implemented by Tryfonidis et al. in 1995. They found that the spatial Fourier technique proved very useful in identifying physical details of modal activity prior to stall, but not suitable for spike stall precursor detection. For modal stall-type compressors, the traveling wave energy analysis worked very well, and was able to predict the onset of modal stall approximately 50 to 100 revolutions before the actual event; however, for the spike stall, the warning period was reduced [54].

Heinlein et al. (2017) [55] used Grubb’s test to track anomalies of entropy of the airflow in time and space, to detect stall precursors. Using this statistical analysis method, they were able to detect stall precursors 16.5 revolutions ahead. However, the application of Grubb’s test in outlier detection requires the number of outliers to be predetermined, which is difficult because the number of outliers that determine the stall condition is not fixed.

Aung and Schoen [56], in 2019, compared several statistical techniques to see which technique gives the best prediction of spike stall. They compared autocorrelation, special entropy, and dynamic model changes as precursor for spike stall events. The dynamic model of the pressure data as an autoregressive linear model was found to predict spike stall by tracking the number of outliers of the eigenvalues of the model. The method was able to predict the spike stall precursor at least 16 rotor revolutions prior to the stall event [48].

Li and Zhang (2019) [57] applied fast wavelets analysis to predict stall. Using this approach, with low frequency reconstruction, modal-type stall was predicted 100 rotor revolutions ahead. The lower frequency reconstruction could predict spike-type stall only 2–3 rotor revolutions ahead. However, using the higher frequency reconstruction, they state that spike-stall could be predicted ahead of time, but they did not explicitly mention that quantitatively.

One of the reasons for spike stall not being able to be predicted earlier is that the spike disturbance grows and develops into a fully developed rotating stall very quickly. According to Vo et al. (2008) [58], spike disturbances form within two to three rotor revolutions. Weichert and Day (2014) [59] also stated that embryonic disturbance grows very quickly, i.e., only three blade pitches of the rotor. Understanding the flow mechanism near the stalling condition is important for identifying and predicting the precursor to spike stall. Numerous simulation and computational studies have been done to understand the flow mechanism near the stalling condition. Wu et al. (2012) [60] did a simulation-based study to explain the flow mechanisms near spike stall inception in an axial compressor rotor. The study identified two primary precursors to spike stall: tip clearance spillage flow and tip clearance backflow. Yamada et al. (2013) [61] also identified the specific flow features at the spike inception, such as tip clearance flow spillage and backflow. Hoying et al. (1999) [62] present some of the other studies investigating the link between spike stall and tip clearance flow. These studies postulate an explanation of the flow mechanism during spike stall inception, including the vortex created at the tip clearance moving toward the leading edge of the blade. However, this may not be sufficient in predicting spike stall to take control action, since spike stall grows into fully rotating stall abruptly after the inception.

Inoue et al. [63] investigate the mechanisms of short and long length-scale stall cells in an axial compressor rotor, focusing on their structural differences and implications for rotor performance under mild and deep stall conditions. Utilizing a low-speed compressor test rig and a double phase-locked averaging technique, they track the pressure and velocity distribution around the rotor. The short length-scale stall cell is characterized by concentrated vortices that resemble a tornado, rotating faster than long length-scale stall cells (LLSCs), and appearing in clusters during mild stall conditions. These vortices cause fluctuating forces on the blades, with higher moment fluctuations than those associated with LLSCs. The study’s insights are significant for predicting compressor stall behavior and the associated risks to blade integrity in practical compressor designs.

Young et al. [64] focus on identifying precursors to stall in axial compressors by analyzing irregularities in blade passing pressure signatures. By investigating the influence of tip clearance size and eccentricity on these irregularities, the authors demonstrate that higher irregularity levels in blade passing signatures correlate with approaching stall conditions, especially in compressors with larger tip clearances. The study emphasizes that irregularities in pre-stall flow are linked to structured disturbances rather than random turbulence, complicating the implementation of stall warning systems in aero-engines where tip clearance varies during operation. This research contributes to the understanding of flow instability behavior, proposing that such irregularities could serve as early warning indicators for stall in practical applications.

Dhingra et al. [65] present a correlation-based method to detect stall and surge precursors by analyzing pressure signals from sensors over the rotor. The study highlights how disruptions in the periodic blade passing signature can indicate approaching aerodynamic instabilities. The researchers find that sensor placement, especially around the rotor mid-chord, is crucial for detecting these disruptions. Additionally, they suggest applications in monitoring blade health and rotor tip clearance. Through tests on low- and high-speed compressors, this approach is validated as a viable precursor to stall onset, providing potential for real-time monitoring systems aimed at enhancing compressor stability and operational safety.

Höss et al. [66] examine the onset of stall in a turbofan engine’s compressor system, investigating how different rotor speeds affect stall precursor types and their detection. Using the LARZAC 04 engine and analyzing pressure fluctuations with Fourier and wavelet techniques, they observe three distinct stall inception mechanisms: spike-type precursors at low speeds, modal wave formations at mid-range speeds, and high-speed stall induced by rotor unbalancing. Notably, spike-type inception dominates when the inlet flow is distorted. The wavelet transform offers the longest pre-warning, detecting stall signs hundreds of rotor revolutions before onset, whereas Fourier analysis provides only brief warning periods. Their findings underscore the value of wavelet analysis in early stall detection, suggesting that active stall avoidance systems could benefit from this advanced predictive capability to improve compressor stability under variable operating conditions.

Tahara et al. [52] propose an innovative stall warning index that relies on correlation decay in pressure signals near the rotor’s leading edge. Testing on a single-stage axial-flow compressor reveals that as the compressor approaches stall, the correlation between successive blade pitch pressures diminishes, offering an early stall warning even before spike inception. The index’s sensitivity to flow coefficient, tip clearance, and rotor blade incidence underscores its adaptability to different compressor designs. The approach has practical advantages: it simplifies real-time monitoring without complex spectral analysis, tolerates minor blade variations, and could allow for a significant reduction in surge margins, enhancing efficiency and lowering operational costs. This work highlights the potential of real-time correlation-based methods to improve stall detection reliability in axial compressors, paving the way for more responsive and efficient control systems.

Chen et al. [67] considered a clean-inlet case to identify the pre-stall behavior of the NASA Stage 35 single stage compressor transitioning into stall using a continuous throttle maneuver with TURBO, a physics-based computational tool. The simulation utilized a full annulus grid to monitor blade dynamics and stall initiation under near-stall conditions. Modal waves emerged as the compressor got to stall following the development of spike, signaling the transition to rotational stall that marks the onset of stall. The dual-shock system was the only observation of stall initiation that intensified the reverse flow region, a critical stall precursor to confirm that stall was driven majorly by vortex–shock interactions.

Si et al. [68] further emphasized how the instabilities arising from rotating stall and surge degrade compressor performance by causing pressure drop and inefficiency, which could sometimes aggravate blade fracture. The Mansoux-C3 model is used in the rapid detection of spike, as it is localized, and detecting it would act as a critical indicator of rotating stall. The simulation utilizes system dynamics identification to record normal and stall inception periods by segmenting them into time intervals. A Radial Basis Function (RBF) network-based estimator, used to approximate and store, is employed to recognize patterns of stall inception with the aid of recognition errors that issue an alarm when the compressor approaches stall. This precedes the onset of rotating stall by 18 revolutions. The investigation confirms that the Mansoux-C3 model can accurately detect the onset of spike-type stalls.

Gong et al. [69] proposed a computational model for a multistage compressor that uses a body force method to approximate blade effects and spike disturbances mimicked by body force impulses. Simulations on a four-stage GE compressor configuration established that among all disturbances, only increasing the amplitude of tip spike disturbances reduced the stall margin, making the compressor vulnerable to instability. The gap between blades was found to be proportional to the stability of the compressor, suggesting prevention of stall in compressor by managing gaps.

Righi et al. [70] studied and validated a 3D simulation tool used for simulating behaviors in a low-speed, three-stage compressor—specifically reverse flow, rotating stall, and surge that displays significant accuracy and efficiency more than traditional CFD approaches. Simulations of the rotating stall gathered the precursor behavior through small flow instabilities, induced due to increasing throttle reduction. The paper found that patterns of distortion at the inlet influenced stall initiation as two types: a more aggressive one that resulted in locked-in stall behavior, and a milder one that caused a stable rotating stall. It confirmed that the triggering mechanism itself affects the stall behavior.

Lou and Key, in 2020 [71], explored the possibility of nonlinear feature extraction such as approximate entropy (ApEn) and Attractor Reconstruction, a time-series data into higher-dimensional phase space converter to capture minor disturbances that occur before deep surge. ApEn relies on four parameters, similarity radius (r), time delay (τ), embedding dimension (m), and data size (N), to pick up spikes on trends that could point to system instability that could foretell stall. The algorithm was implemented in two cases, a Single-Stage High-Speed Centrifugal Compressor that exhibited minor surge during deceleration, and a Three-Stage Axial Compressor that supported the validating of ApEN in prediction of stall, as it aligned with the peaks of nonlinear disturbances across various throttle settings.

6. Upcoming Methods

Artificial intelligence (AI) techniques have emerged as powerful tools for detecting and predicting stall and surge precursors in axial compressors, addressing a critical challenge in turbomachinery operation. This section examines recent peer-reviewed research on AI-based approaches for stall and surge detection, focusing on modeling, precursor identification, and control strategies.

Machine learning algorithms have shown significant promise in detecting stall and surge precursors. Various supervised learning techniques have been applied to classify compressor operating conditions based on sensor data. Ying et al. (2020) [72] proposed a novel compressor performance modeling method based on a support vector machine (SVM) nonlinear regression algorithm. The approach demonstrated superior interpolation and extrapolation performance compared to traditional neural network methods such as back propagation (BP), radial basis function (RBF), and Elman neural networks. The SVM-based method achieved root mean square errors (RMSE) of 2.72% for flow characteristic map representation and 1.81% for efficiency characteristic map representation, outperforming other neural network algorithms by up to 47%.

Deep learning, particularly Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs), has shown success in processing complex time-series data from compressors. Liu et al. (2024) [73] utilized Long Short-Term Memory (LSTM) networks, a type of RNN, to predict stall initiation in a low-speed axial contra-rotating compressor. The study demonstrated that LSTM models could provide early warnings for stall in 11 out of 15 speed configurations. To address the limitations of LSTM in certain scenarios, the work developed a hybrid CNN-LSTM model that showed improved performance, successfully issuing warnings before stall occurrence for all speed configurations.

AI-driven stall and surge detection heavily relies on high-quality data from various sensors. Recent studies have focused on optimizing sensor placement and data preprocessing techniques to enhance the performance of AI models. Jiang et al. (2019) [74] addressed the challenge of limited sample data in constructing performance maps for centrifugal compressors. The study developed a loss-analysis-based model that could provide accurate predictions even with a small dataset, outperforming traditional interpolation and neural network methods in certain scenarios.

Recent work by Zanotti et al. (2024) [75] proposes an AI-based method for detecting the onset and type of compressor instability using dynamic model parameter estimation. The proposed approach leverages artificial neural networks (ANNs) to identify both rotating stalls and surges with high accuracy. This study demonstrates the potential of AI in enhancing the fidelity of compressor instability models, potentially improving upon traditional Moore–Greitzer-based approaches. The methodology involved following key components:

• Using the Moore–Greitzer model to generate time-series data for different compressor operating conditions (stable, rotating stall, deep surge, and mixed instability).
• Developing an artificial neural network (ANN) to estimate the model parameters (Greitzer’s B parameter and throttle setting) from the generated time-series.
• Training the ANN using TensorFlow 2.0 with different input data configurations, including full and reduced time-series lengths.
• Evaluating the ANN’s performance in detecting instabilities under various conditions, including missing data and sensor inaccuracies.

The authors found that their AI-based approach could accurately estimate the Moore–Greitzer model parameters and distinguish between different types of instabilities. The method showed robustness even with reduced input data and sensor inaccuracies, demonstrating its potential for real-time instability detection in compressors.

While AI techniques have shown promising results in laboratory settings, their implementation in real-time industrial environments poses several challenges. Ying et al. (2020) [72] highlighted the importance of real-time performance in the comparison of different AI algorithms. The SVM-based method demonstrated better real-time performance compared to traditional neural network approaches, making it more suitable for on-site, real-time applications.

Machine learning techniques have shown promise in the early detection of compressor stall precursors. Ames Laboratory and NETL researchers (2020) [76] explored the use of LSTM networks for predicting compressor stall. The study utilized real compressor stall datasets from a 100 kW recuperated gas turbine system. The LSTM model, particularly in its regression configuration, demonstrated the ability to predict stalls 5–20 ms before occurrence, providing a crucial window for preventive action. Another study by Jin et al. (2023) [77] introduced an anomaly detection method, utilizing DeepESVDD (deep ellipsoid support vector data description). This method analyzes dynamic pressure signals to identify early signs of stalling. The resulting warning indicators are then fed into a CNN to develop a classification model for rotating stall warning in aero-engine compressors. By combining machine learning methods with anomaly detection techniques, the work addresses the challenge of identifying early indicators of rotating stalls, which is critical for maintaining engine performance and safety.

In addition, Ref. [78] discussed surge precursor detection utilizing a hybrid network and spatiotemporal features. The method involves collecting pressure data from a compressor’s blade passage while operating a single-stage axial compressor near the optimal point until a spike stall is observed. The study emphasized using graphs to capture relations among clusters of pressure transducers, while nonlinear dynamic transient characteristics are captured from the sequence of graphs and RNN combined. A validation accuracy of 93–100% for precursor detection at 30 revolutions prior to stall initiation is reported. The study exercised data generation strategies based on symmetric placements of multiple sensor-chords due to a lack of experimental data volume. The author suggested the prospect of the approach by including larger experimental data for model training.

In Ref. [79], a compressor modeling method utilizing deep learning is proposed. The study describes an axial compressor characteristic map, which is modeled by an LSTM, and a single-stage axial compressor test stand rig, which is modeled with the Toolbox for the Modeling and Analysis of Thermodynamic Systems (T-MATS) to serve as the ground truth. The method involves segmenting the characteristics map to tiles of multiple nonlinear MIMO sub-systems with input being the throttle and rotor RPM, and output being the mass flow coefficients and pressure coefficient. The study proposed the construction of a nonlinear MIMO model based on LSTM presenting each of the tiles. The study suggests promising results from the LSTM-based models to mimic the ground truth for varying conditions such as RPM and throttle opening.

Despite these advancements, the field still faces significant challenges. While compressor surge and stall have been extensively researched, no methods for early prediction have been proven universally effective. This highlights the complexity of the problem and the need for continued research and validation of AI-based approaches across different compressor designs and operating conditions.

Researchers have also explored hybrid models that combine different AI techniques or integrate AI with physics-based models. Elhawary et al. (2023) [80] proposed a combination of machine learning and genetic algorithms to optimize air jet parameters for controlling rotating stall in an axial compressor. The study used shallow neural networks to model the influence of air jet parameters on surge margin improvement and power balance, followed by employing a genetic algorithm to optimize these parameters for different rotational velocities.

In Ref. [81], an axial compressor airfoil design is proposed, utilizing a neural network to address the essential coherence between geometric design parameters and the aerodynamic model criteria. The model parameters are further optimized to confirm the global optimum solution utilizing a genetic algorithm (GA).

Intelligent control applied to compressor systems can be extended to identifying the operability of the compressor from the combination of domain knowledge and a trainable model that infers the system conditions in real-time. Taylor et al. (2020) [82] presents a novel methodology called MRP (Machine learning, Rapid testing, and Physical parameterization) for predicting the operability of damaged compressors. The method combines machine learning techniques with rapid experimental testing capabilities and physics-based parameterization of blade damage. By using neural networks trained on data from 125 different damage configurations tested rapidly, and incorporating 10 physically meaningful parameters to describe the damage, the authors were able to predict compressor stall points with 2% accuracy. The results show that this approach can predict operability more accurately than human experts, while also generating new physical insights into how different types of blade damage affect compressor performance. Key findings include the importance of gaps between damaged blade clusters and the nonlinear relationship between damage magnitude and operability. The authors conclude that this methodology has broad potential for tackling complex aerospace problems by augmenting human understanding with machine learning capabilities.

Lou et al. (2020) [71] introduce an approach for detecting small nonlinear disturbances prior to compressor stall using nonlinear feature extraction algorithms, specifically phase reconstruction of time-series signals and evaluation of approximate entropy. This method is applied to data from two different compressors: a high-speed centrifugal compressor and a multistage axial compressor. In both cases, spikes in approximate entropy are observed prior to the surge, indicating the presence of nonlinear disturbances. For the centrifugal compressor, these spikes occurred 8–10 s before the fully developed stall, potentially providing enough warning time to avoid stall. For the axial compressor, pre-stall disturbances are detected using approximate entropy for both modal- and spike-type stall inception. The results demonstrate the potential of using approximate entropy as a parameter for small disturbance detection and stall warning across different compressor types and stall inception mechanisms.

A study on using artificial intelligence to generate and optimize axial compressor performance maps is presented in Ref. [83]. The author proposed an automated workflow to generate training data for multistage axial compressors with variable inlet and stator vanes. A neural network is trained to predict compressor performance with less than 3% error. Key results reported include training of the neural networks to predict the performance of three different compressor designs. Utilizing the trained models to optimize variable vane angles along an operational line achieved pressure ratio predictions within 2% error, and efficiency predictions within 1% error compared to physics-based solvers. The work demonstrates that the AI approach can evaluate optimal vane angles and compressor efficiency much faster than traditional methods when analyzing many operational points.

Shimin et al. [84] describe a new method for early detection of compressor stall and surge in aircraft engines, called Multiscale CNN-SVM-FC. The key results show that the method, which combines multiscale detection windows with a Convolutional Neural Network-Support Vector Machine (CNN-SVM) algorithm, achieved over 99% accuracy in identifying unstable states when tested on data from a five-stage axial compressor at various speeds. To minimize incorrect alerts, the system combines two approaches: it uses a fuzzy control algorithm and aggregates predictions over time from the multi-branch network. These methods work together to perform a comprehensive analysis, ultimately improving the accuracy of the results. Compared to traditional stall prewarning methods, it provided warning signals an average of 164 milliseconds earlier, while also reducing false alarm rates compared to standard CNN-SVM models. Additionally, the method reduced the uncertainty associated with threshold selection, which is typically based on engineering experience in traditional approaches.

Rauseo et al. [85] discussed the development of an approach to predict flutter in aircraft engine fans and compressor blades using a combination of machine learning techniques and reduced order models. The research [85] addresses the growing need for fast and accurate prediction tools in modern aircraft engine design, which faces increasingly complex requirements for performance, cost, emissions, and noise reduction. The study’s key innovation lies in its physics-guided framework, which incorporates prior knowledge of flutter by formulating relevant steady-state input features and integrating results from low-fidelity analytical models. The results are particularly promising, demonstrating that the developed models can accurately predict flutter stability for unseen cascades, even when trained on a single geometry. Importantly, these models allow for flutter prediction based solely on steady-state flow, eliminating the need for computationally intensive unsteady simulations. This work could enable rapid assessment of blade flutter stability under various mechanical properties at minimal additional computational cost once the mean flow is known, potentially contributing towards the efficiency of aeroelastic stability analysis in aircraft engine design.

While the paper is primarily focused on precursor detection and prediction in an axial flow compressor system, the implications for control strategies are significant. The ability to predict stall or surge events with a 5–20 ms lead time, as demonstrated in the LSTM study [76], opens up possibilities for implementing high-speed control interventions to prevent instability onset.

Overall, AI techniques have demonstrated significant potential in improving the detection and prediction of stall and surge precursors in axial flow compressors. From machine learning approaches like SVM to deep learning models such as LSTM and CNN, these methods offer improved accuracy and real-time performance compared to traditional techniques. However, challenges remain in terms of data availability, model interpretability, and real-time implementation. As research in this field progresses, it is expected to see more robust and efficient AI-driven solutions for maintaining the stability and performance of axial flow compressors.

7. Conclusions

Compressors are complicated systems that have limiting issues such as stall and surge. These characteristics have been a focal point of many studies resulting in numerous design propositions and modifications for axial compressor system for more than 80 years. As these issues are closely related to safety and efficiency, a deeper understanding of the stall phenomena has been sought. This quest has led to models that capture the general dynamics of axial compressor systems, as well as to models of specific phenomena that are believed to be partially responsible for the stall event. For example, a number of theories have been proposed on the basis of the relationship between the tip clearance and the associated tip clearance leakage flow with the ensuing stall cells in the blade passage, the blade failures, and the development of part-span as well as full-span stall. Additionally, researchers have proposed maps that chart the progression of stall from stall precursors, stall development, to stall events based on the momentum of the tip leakage flow and the tip gap.

Controlling the onset of stall, or even inducing flows that mitigate the onset of the stall phenomena have been proposed on the basis of the gained understanding of the underlying physics. One of the passive control methods investigated is the use of grooves in the casing of the compressor. Researchers studied the geometry of these grooves and the effect the resulting flow disturbance has on the tip leakage flow, and hence on the development of stall events. For active control measures, a prediction of stall is required to allow for sufficient time to intervene and prevent the development of stall. Based on the geometry of the compressor, stall initiation and development of stall manifest themselves as either a modal wave or a spike in pressure. Hence, the type of stall is referred to either modal stall or spike stall. For stall precursor detection of either type of stall, researchers have proposed data analysis tools based on the monitoring of the pressure at the casing of the first compressor stage. These tools include the use of statistical methods, time-frequency analysis, and time-series analysis methods. In the case of modal stall, some of the proposed methods allow for a prediction of the stall event several hundred rotations prior to stall, while for spike stall, the available methods have only been able to predict the onset of stall by single-digit or double-digit revolutions prior to the stall event. For spike stall precursor detection, some of the models proposed capture the dynamics of the flow within a blade passage based on the measurement of casing pressure. Recent work also utilizes machine learning, including deep learning methods to predict stall precursors.

Active control of these initiation events—and to some degree the development of stall—often involves the injection of air to perturb the flow in the tip region of the rotor. Numerous control algorithms have been tested in laboratory settings, where pressurized air from later stages of the axial compressor were tapped and fed to the stage with active stall control. As one of the motivations for stall control is the improved efficiency of the compressor, these active control methods are not without compromise, and safety concerns will impose rigorous testing and analysis requirements before such systems are deployed in commercial settings.

The following

Table 1. Summary of methods and outcomes by section and topic.

Modal and Spike Stall
Category	Approach	Performance	Year	Ref
Numerical Simulation	A Series of nonlinear equations to capture pressure rise, flow coefficient variations, and rotating stall cells during transients using a Galerkin approach.	Has a become a baseline for compressor models. Highlighted that rotating stall is generally quasi-steady and non-axisymmetric, while surge remains unsteady and axisymmetric. Valid for both constant and variable speed conditions.	1986	[7]
Numerical Simulation	Extended previous work by Moore and Greitzer.	Adjusted the set of equations developed by Moore and Greitzer to include a fourth state. Model is able to show both stall and surge.	1997	[12]
Experimentation	Tested previous found numerical model using experimentation.	Showed that the numerical model found by authors was able to be used for a control scheme capable of accurate control.	1997	[13]
Numerical and Data Driven Model	Adaptive neuro-fuzzy inference system to predict compressor performance across different speeds.	Limited datasets are used to build comprehensive performance sets. Noted that surge and rotating stall were directly influenced by load line positioning and speed. Showed larger plenum volumes increased instability. Shows how different parameters affect performance.	2010	[14]
Data Driven Model	Spatiotemporal visualization techniques.	The use of computational tools and visual analytics were utilized to provide more efficient detection of rotating stall through the tracking of vortex dynamics.	2016	[15]
Numerical Simulation	Data mining techniques used to identify vortex and employed the Line Integral Convolution (LIC) method for near stall conditions.	Showed the connection between stall propagation and complex vortex formation. Gives a deeper understanding of the mechanics behind flow separation and stall inception.	2017	[16]
Tip Clearance and Tip Leakage
Category	Approach	Performance	Year	Ref
Experimental	Experimentation performed to determine how tip clearance and blade shrouding affects stall.	1% of blade height clearance inhibits local stalling at tip. Smaller clearance leads to an increase in stall. Shrouded blading clearance of 1% blade height decreases performance.	1958	[22]
Data Driven Model	Tip clearance changes analyzed to determine rotor performance and then compared with flow model.	Found that larger tip clearance increases tip loading and that when operating below designed flow rate, tip stall develops.	1960	[23]
Literature Review	A survey of tip clearance.	Analyzed flow in boundary layer and efficiency. Found rotating blades can reduce efficiency when scraping of boundary layer occurs.	1968	[21]
Numerical Simulation	Utilized a numerical model at a range of tip chord ratios to observe onset of stall, flow range, and stability.	Found that increasing the tip clearance increased onset of stall. Zero tip clearance showed zero leakage vortex, while non-zero tip clearance has more instability due to leakage vortex and flow blockage.	1993	[17]
Numerical Simulation	Computational fluid dynamics is used to identify spike disturbances.	Both the growth of trailing-edge backflow and leading-edge spillage are required for spike disturbances.	2001	[24]
Experimental	Experiment uses inlet guide vanes to adjust blade height to determine the connection between radial loading and stall inception.	High loading distribution near the stall indicates spike stall while changing conditions near the hub do not affect stall inception.	2003	[25]
Data Driven Model	Used three types of fluidic actuators on the casing wall to look at effectiveness of tip clearance flow control.	Four criteria were found, including actuator, pitchwise location, amplitude of actuation, and frequency. Streamwise momentum injection showed improved tip clearance control.	2005	[19]
Numerical Simulation	Experimentation with computational fluid dynamics is used to tip leakage.	Found that the spike disturbances and rotating inlet distortion are related by how often the two cross paths.	2007	[26]
Numerical Simulation	To verify previous experimentation, simulations were done to find unsteady behavior at tip flow at near-stall conditions.	Tip secondary vortex (TSV) was found to move, causing a hump in the casing pressure measurements. This TSV is caused by a breakdown of tip leakage vortex.	2012	[27]
Numerical Simulation	A Fast Fourier Transform is used to evaluate previous experimentation.	Able to replicate experimental results that show a TSV that is found to be a critical cause of instability.	2014	[28]
Experimental	Used experimentation to methodically vary chord tip clearance, angle of incident, and Reynolds number. Utilized a water channel and inked fluid elements to observe tip leakage.	Instability appears between an angle of incident of 17.5° and 20° under stable conditions. Unstable conditions differ from stable.	2017	[20]
Experimental	Experiments ran on 3-stage axial compressor running at an intermediate speed. Different tip clearances ranging from 1.5 to 4% span were tested.	Results showed the opposite of previous studies that indicated the spike stall goes from short-scale to long-scale with increased tip clearance, whereas this study found that increasing tip clearance transitioned from modal to spike.	2017	[29]
Experimental/Simulation	An experiment along with computational fluid dynamics is used to analyze tip clearance in a single-stage axial compressor.	Found instability to progress from stable to rotating instability to rotating stall. Tip leading vortices were found to disturb the pressure, and that closer to stall, scatter vortices develop.	2020	[30]
Physical aspects that affect stall
Category	Approach	Performance	Year	Ref
Porous Casing	Experimentation performed using porous casing to determine performance.	Found improvement in stability with porous casing, but some reduction in performance	1970	[34]
Porous Casing	Experimentation performed using porous casing to determine performance.	Found that stall-margins increased with porosity, and efficiencies were higher than with a solid casing	1971	[35]
Grooved Casing	Experimentation performed using porous casing to determine performance.	Circumferential groove, axial-skewed, and blade angle were tested, with the results showing that all improved stall with a slight loss in efficiency	1974	[36]
Tip Injection	Experimentation is performed to determine how air injection upstream of the first rotor blade affects performance.	Found that injection improves stall margin by more than 13% with no effect on efficiency. In particular, they studied self-recirculating injection, as it proved to be the most efficient.	2018	[50]
Tip Leakage Flow	Simulation with experimentation to determine cause of tip leakage blockage.	Found that the cause of blockage at zero clearance is corner separation and that as clearance increased, a clearance maximum flow range is achieved	2019	[47]
Grooved Casing	A numerical study is performed to identify blockage parameter.	Peak blockage is found to be at 10% of the tip chord aft of the tip edge. An optimized groove is then found to improve stall margin.	2021	[48]
Grooved Casing	Numerical studies and experimentation is performed to determine how axial skewed slots affect instability.	Found an 8% improvement in stall margin with a decrease in the frequency broadband hump.	2021	[49]
Stall Precursor
Category	Approach	Performance	Year	Ref
Experimentation	Experimentation performed to find precursors of rotating stall.	Found that rotating stall occurs after a pressure fluctuation periodicity collapse, and a parameter that represents this periodicity is found.	1991	[51]
Experimentation	Nine compressors are experimentally analyzed to find stall and surge precursors.	All compressors show small amplitude waves that travel around the circumference of the casing before stall occurs. At about 0.5 shaft speed for low-speed compressors, these waves develop hundreds of revolutions before stall.	1995	[53]
Experimentation	Experimentation performed to relate physical phenomena with stall control systems.	Found that low-speed compressors have spike stall, while mid-speed have modal stall. Found a new nonrotating stall in three of the four compressors analyzed. With the range of physical phenomena, it was deemed that an active control system would not be ready for practical use in the near future.	1999	[54]
Computational Analysis	Computational analysis is performed to analyze three-dimensional flow structures and how they relate to stall.	Simulated a short length-scale spike inception that showed the tip clearance vortex moving forward of the leading edge.	1999	[62]
Computational Analysis	A computational model is used to simulate stall inception.	The developed model found that rotating stall developed due to short-wavelength disturbances and that switching from long- to short-wavelengths is a result of re-staggering inlet guide vanes. They defined where disturbances occurred with short-wavelength disturbances occurring in the rotor blade row. This strength increased within the stators. To reduce this growth, reducing inter-blade row gaps is recommended.	1999	[69]
Experimentation	Various techniques are used to measure stall inception on a two-spool turbofan engine.	Different analyzing techniques are used including temporal low-pass and band-pass filtering, Fourier transforms, and wavelet analyzing technique. In the low-pressure compressor, three types of processes for stall inception were observed. At low-speed, spike-type precursors occurred prior to stall. At mid-speed, long wavy pressure fluctuations occur. For inception detection, wavelet transform predicted stall a few hundred rotor revolutions in advance.	2000	[66]
Experimentation	Experimentation performed to compare short length-scale stall cells to long length-scale cells.	Found that short length-scale stall cells for a vertex ahead of the rotor blade that spans from the casing wall. Long length-scale stall cells have a vortex that separate in the front half of the cell to the center and re-enter at the rear half of the rotor on the hub side.	2001	[63]
Experimentation	A correlation scheme with experimentation is used to describe a new technique for stall and surge precursor detection.	A momentary lapse is detected at the onset of instabilities that can be used for precursor detection. This method was implemented on both high-speed and low-speed compressors. It is found that the location of the pressure sensors is important, and mid-chord on the rotor is determined to be the best location.	2003	[65]
Experimentation	Experimental analysis is performed to find stall precursors using high-response pressure transducers on the rotor’s leading edge.	A risk index is developed based on the correlation between degrading pressure histories of current. This index is affected by many factors that need more study for practical use.	2007	[52]
Simulation	Simulations are used to show how the dynamics of tip clearance flow affect stall.	Found that spike stall disturbances begin with the initiation of backflow at the trailing-edge of the tip clearance and parallel flow to the leading-edge of the tip clearance.	2008	[58]
Computational Analysis	Computational fluid dynamics is used to simulate flow prior to stall in a transonic compressor.	Rotating stall is shown in simulation to begin from instabilities. A rotating long-length disturbance that is followed by a spike-type breakdown. Long-length waves are attributed to spike inception.	2008	[67]
Experimentation	Experimentation is performed to determine internal flow fields in relation to spike-type stall.	Marked to disturbances in flow that are designated as B1 and B2. Tip secondary vortices (TSV) are the instigator for these disturbances. B2 is located at the tip-front, which made the TSVs a stronger instigator for spike-type stall.	2012	[60]
Simulation	Numerical simulation performed to analyze unsteady flow in multistage axial compressor ran near-stall.	Using a computational mesh with detached-eddy simulation (DES) and data mining of vortex identification and streamline drawing, it was found that corner separation on the hub side was related to stall.	2013	[61]
Experimentation	Experimentation is performed to characterize the irregularities that cause stall.	It is found that irregularities are dependent on tip clearance size and eccentricity. A tip clearance that is small and uniform will only have a modest reduction in flow rate, while an enlarged tip clearance will have a sharp rise in irregularity. Additionally, it was found that irregularities in pre-stall flow are not random, but a coherent flow structure.	2013	[64]
Experimentation	Experimentation is performed to determine new measurements for spike-type stall inception.	Results show an embryonic disturbance that leads to a clear spike. This begins as a small disturbance in the blade passage and that this can only be seen upstream once it has increased in size.	2014	[59]
Computational Analysis	A deterministic learning method is used to model stall precursors for rapid detection.	Through the rapid detection of small oscillation faults, spike-type precursors are detected. This shows that deterministic learning is a viable method to detect rotating stall.	2016	[68]
Statistical Analysis	A statistical analysis with Grubbs’ test is used to detect anomalies and trends.	Found that a rotating stall cell formed during rotating disturbance region. Additionally, found that a spiral-type vortex appeared at the tip of the clearance vortex. These two behaviors caused the tip vortex oscillation to increase radially and circumferentially leading to rotating stall.	2017	[55]
Computational Analysis	Cylindrical Euler equations are used to model three-dimensional through-flow.	Uses the Godunov solver to model three-dimensional rotating stall and surge to accurately calculate inter-cell fluxes. Model validated on low-speed three-stage axial compressor. Found that instability development affected by tank volume and level of distortion.	2018	[70]
Experimentation	Evaluated different precursor identification methods to determine their ability to be used for precursor detection.	Methods analyzed include physical measurements, outlier detectors, entropy, and Autoregressive (AR) models. The study found that the best method is an AR with Generalized Extreme Studentized Deviate Test (ESD).	2019	[56]
Computational Analysis	Wavelet tool is used to predict stall precursors for both spike and model inception in an experimental setup.	Found that modal stall can be predicted using low frequency wave reconstruction 100 revolutions before stall. For spike stall, low frequency is not adequate. High-frequency reconstruction can be used with a frequency band of 0.2–0.8.	2019	[57]
Computational Analysis	Nonlinear feature extraction algorithms are used to evaluate a parameter called approximate entropy.	Stall datasets are used from two different compressors. The parameter approximate entropy spike prior to surge in both compressors. Approximate entropy can then be used as a stall warning.	2020	[71]
Upcoming Methods
Category	Approach	Performance	Year	Ref
Compressor modeling	Design of axial compressor airfoils with artificial neural networks and genetic algorithms.	Essential coherence between geometric design parameters and the aerodynamic criteria are identified using a simple neural network.	2000	[81]
Compressor modeling	Performance prediction of the centrifugal compressor based on a limited number of sample data.	The study indicates that for predictions within the data range, the loss-analysis-based models yield more accurate forecasts, even with limited data. These models also demonstrate consistent performance. In contrast, the neural network model requires a larger dataset with additional speed lines to produce superior results.	2019	[74]
Compressor modeling	Compressor performance modeling method based on support vector machine nonlinear regression algorithm.	The proposed method compared to the three other algorithms including Radial Basis Function, Elman Neural Network, and Backpropagation, showed better real-time performance.	2020	[72]
Stall Precursor detection	LSTM classification/regression on stall dataset.	Stall precursor identification 5–20 ms in advance of the stall activity.	2020	[76]
Intelligent Control	Machine learning, Rapid testing, and Physical parameterization for predicting the operability of damaged compressors.	This approach can predict operability more accurately than human experts, while also generating new physical insights into how different types of blade damage affect compressor performance.	2020	[82]
Stall Precursor detection	Compressor stall warning using nonlinear feature extraction algorithms.	Spikes in approximate entropy are observed prior to surge, indicating the presence of nonlinear disturbances.	2020	[71]
Compressor modeling	Axial compressor map generation leveraging autonomous self-training artificial intelligence.	AI approach can evaluate optimal vane angles and compressor efficiency much faster than traditional methods when analyzing many operational points.	2023	[83]
Stall Precursor detection	Anomaly detection method, utilizing DeepESVDD.	The author claimed the experimental results showed significant accuracy in the stall precursor classification.	2023	[77]
Intelligent Control	Machine learning and genetic algorithms to optimize air jet parameters.	Global optimal parameter achieved for velocity ratio.	2023	[80]
Compressor modeling	Modeling axial compressor systems using deep learning methods.	The study suggests results from the LSTM-based models to mimic the ground truth for varying conditions such as RPM and throttle opening.	2023	[79]
Stall Precursor detection	Precursor detection utilizing a hybrid network and spatiotemporal features.	A validation accuracy of 93–100% for precursor detection at 30 revolutions prior to stall initiation is reported.	2024	[78]
Stall Precursor detection	Long Short-Term Memory (LSTM) networks to predict stall initiation in a low-speed axial contra-rotating compressor.	Hybrid CNN-LSTM model showed improved performance, issuing warnings before stall occurrence for variable speed configurations.	2024	[73]
Stall Precursor detection	AI-based detection of surge and rotating stall in axial compressors via dynamic model parameter estimation.	Developed an artificial neural network (ANN) to estimate the model parameters, Greitzer’s B parameter, and throttle setting from the generated time-series data.	2024	[75]
Compressor modeling	Prediction of flutter in aircraft engine fan and compressor blades using a combination of machine learning techniques and reduced-order models.	The results are particularly promising, demonstrating that the developed models can accurately predict flutter stability for unseen cascades, even when trained on a single geometry.	2024	[85]
Stall Precursor detection	The method combines multiscale detection windows with a Convolutional Neural Network-Support Vector Machine (CNN-SVM) algorithm.	Compared to traditional stall prewarning methods, it provided warning signals an average of 164 milliseconds earlier, while also reducing false alarm rates compared to standard CNN-SVM models.	2024	[84]

summarize the different contributions discussed in this review by section and topic. The tables include the individual references and a corresponding summarizing statement of the work presented in the paper. As the research of stall in axial compressors spans almost eight decades, this review paper attempts to capture the main contributions, knowing that many interesting and noteworthy works have been left out for brevity.