Modeling And Rapid Detection Of Rotating Stall In Axial Compressor Via Deterministic Learning

Aero-engine is known as the “heart” of aircraft. Without high-performance aircraftengines, there would be no advanced aviation weaponry and civil airplanes with power-ful market competitiveness. Turbofan (referred to as the vortex fan) engine is the mostcommonly used power unit of the military and large civil aircraft in the world. Themost notable feature of turbofan engine is that it works with high efciency at highsubsonic/supersonic fight conditions. Since the development of the aero-engine usuallytakes about15-20years, countries that lead in aero-engine technologies in the world at-tach great importance to the relevant basic and applied basic research. High-performanceaero-engine is an important symbol of a nation’s industrial level. Major developed coun-tries in the world, always take aero-engine industry as the priority industry, and attachgreat importance to the associated basic research. Our country is one of the few countrieswith independent capability of aero-engine development in the world. However, comparedwith the world’s advanced level, there are still obvious gaps in terms of reliability, stabil-ity and efciency. Therefore, it is urgent for China to increase investment and acceleratethe development in the feld of aero-engine. Rotating stall and surge are important andchallenging problems in the area of axial compressors. For the reasons that, rotatingstall precedes surge in many machines, rotating stall has received much more attentionin both experimental and theoretical studies. To implement active control of rotatingstall and surge, it is essential to achieve accurate modeling and rapid detection of stallprecursors. Based on its strong points and importance both in theoretical research andpractical application, stall inception detection will be further studied based on the de-terministic learning theory in this dissertation. The main contribution and innovation ofthis dissertation are summarized as follows:1. A precursor for Pitchfork bifurcation in axial compression system was proposed.Firstly, the bifurcation behavior of Moore-Greitzer model was analyzed: A Pitchforkbifurcation in this model is relevant to rotating stall; A Hoph bifurcation is associatedwith surge. The former bifurcation comes up before the latter one. Consequently, aprecursor for Pitchfork bifurcation can be treated as rotating stall. Secondly, based onthe bifurcation behavior of Moore-Greitzer model, a precursor for Pitchfork bifurcationwas proposed via deterministic learning, which was recently presented to learn unknownnonlinear system dynamics from uncertain dynamic environments. Specifcally:(i) severaltypical patterns in Moore-Greitzer model were identifed by deterministic learning, the obtained knowledge of the approximated system dynamics is stored in constant RBFnetworks;(ii) A bank of estimators are constructed using the constant RBF networksto represent the training patterns and previously learned system dynamics is embeddedin the estimators;(iii) By comparing the set of estimators with the test pattern, a setof recognition errors are generated, and the average L1norms of the errors are taken asthe similarity measure between the dynamics of the training patterns and the dynamicsof the test pattern. Therefore, the test pattern (Pitchfork bifurcation) similar to one ofthe training patterns can be rapidly recognized according to the smallest error principle.Simulations results illustrate the approach.2. We present an approach for approximately accurate modeling and rapid detectionof stall precursors. Firstly, a method of modeling the system dynamics corresponding tostall precursor is presented:(i) The Mansoux model, which is a high-dimensional ODEmodel used to approximate the axial compressor, is considered as the approximation ofthe model describing stall precursors.(ii) By analyzing the properties of the Mansouxmodel, for a measurement point, the system dynamics represented by states at this pointand other two adjacent points are supposed to be unknown. The other system dynamicsare assumed to be known.(iii) By using RBF neural networks (NN) and the deterministiclearning algorithms, approximately-accurate modeling of the dominant system dynamicscorresponding to stall precursors is achieved. The obtained knowledge of system dynamicsis stored in constant RBF networks, and is considered as the locally accurate approxima-tion of the model describing stall precursors. Secondly, a scheme for rapid detection ofa stall precursor is proposed:(i) By using the constant RBF networks obtained above,a bank of estimators are constructed corresponding to trained stall precursors.(ii) Bycomparing the set of estimators with the test monitored system, a set of residuals aregenerated.(iii) Based on dynamical pattern recognition, the occurrence of stall precursorcan be rapidly detected according to the smallest residual principle. Simulation studiesare included to show the efectiveness of the approach.3. A low-speed axial fow compressor test rig of Beijing University of Aeronau-tics and Astronautics is employed to verify the efectiveness of the proposed detectionmethod. Five ftting seats are evenly placed in the circumferential direction of the com-pressor. Respectively, fve high response transducers, which are used for dynamic pressuremeasurement for the fow feld between the rotor and stator, are spaced at each fttingseats. The data sampled by the pressure transducers located was saved as voltage values.The voltage values are transformed to axial velocity coefcients by linear transformation.Firstly, by employing the dynamic pattern recognition algorithm proposed in this disser- tation, the data acquired are processed of-line. The general process for the of-line dataprocessing consists of two phases: the identifcation phase and the recognition phase. Inthe training phase, locally-accurate identifcation of the stall inception system dynamicsis achieved by using radial basis function (RBF) neural networks (NNs) through deter-ministic learning. The obtained knowledge of the approximated gait system dynamicsis stored in constant RBF networks. A bank of estimators are constructed using con-stant RBF networks to represent the training stall inception patterns. In the recognitionphase, by comparing the set of estimators with the test stall inception pattern, a set ofrecognition errors are generated, and the average L1norms of the errors are taken as thesimilarity measure between the dynamics of the training stall inception patterns and thedynamics of the test stall inception pattern. Therefore, the test stall inception patternsimilar to one of the training gait patterns can be rapidly recognized according to thesmallest error principle. Secondly, the online programs for stall inception detection arerealized by LabVIEW. Sufcient online experiments were carried out to demonstrate theefciency of the algorithm proposed in this dissertation. Experimental results shows thepossible value of the stall inception detection method in engineering. In this dissertation,information about fow rate and pressure rise in the compressor is employed to identifythe internal system dynamics corresponding to stall inception. Compared to other de-tection methods in the literature, much more information are exploited to achieve moreaccurate detection results.