The Realization Of Turbofan Engine Rotating Stall Detection System Based On GPU And MATLAB

In recent years, turbofan engines has been most commonly used among the world military and civil aircrafts, axial flow compressor which is the core part of a turbofan engine, whether its operating conditions is normal or not has crucial impact on the performance of turbofan engine. During operation of the axial flow compressor, the airflow fluctuations and impact may cause instability of axial flow compressor, and even leads to turbofan engine stop, causing the crash, casualties and other serious consequences. During operation of axial flow compressor, two common system instable phenomenon are rotating stall and surge, and there is always a rotating stall occurs before the appearance of the surge。Therefore rotating stall is considered as a harbinger of the surge. If in the initial stages of rotating stall occurs, we on timely monitoring its signal and gives early warning and regulation, then the reliability and stability of the turbofan engine will be greatly improved. At present, China has independently developed the aero engine with no function to prevent the surge, which is a serious constraint for the performance of the aircraft.Determine learning theory is developed in recent years, which is new theoretical results on the machine learning in dynamic environment. Through the use of powerful learning capacity of RBF neural network, we have access to express, store and use knowledge in unknown dynamic environment. Dynamic pattern recognition is based on the determining learning theory, relates to description of dynamic pattern, definition of similarity and rapid identification of the unknown dynamic pattern. Using determine learning theory and dynamic pattern recognition method can realize system dynamics of the axial flow compressor accurately modeling and rapid identification of the faults, providing a new means to solve the problem of axial flow compressor rotating stall detection. By using RBF neural network to learn dynamic process of rotating stall and save constant weights as a pattern. When the monitoring system is running, it will continue comparing testing data with existing stall pattern, give the appropriate warning signs in according to the matching degree between the two, so as to achieve the purpose of early detection of rotating stall.In practical engineering applications, the calculate amount with the use of determine learning theory and dynamic pattern recognition algorithm in rotating stall detection is relatively large, mainly in the weight calculation of RBF neural networks and identify testing data using multiple patterns simultaneously. To speed up the running of the system, this paper achieves a fine-grained parallel decomposition algorithm with MATLAB parallel toolbox and GPU massively parallel computing power. Parallel transplantation of algorithm is based on software environment combining MATLAB and Jacket Engine. The actual test proved that the algorithm become economical and efficient after parallelization, this method can save both time and space, making the whole computer resources fully utilized. The running speed of system has improved significantly, and the performance is stable, having a good ability to adapt for increase in the size of the neural network and the number of recognition pattern. Algorithm parallelization is mainly using matrix operations and large-scale multi-threaded GPU-based computing, dependence on platform is weak and it is very easy to transplant, so it has good application value. On the basis of achieving accelerated computing, this paper using MATLAB GUI subsystem built a axial flow compressor rotating stall offline detection and analysis system, which is mainly used for offline analysis of the data, including data processing, state analysis, parameter setting, learning and identification, analysis results, capable of offline learning, modeling, recognition and analysis for axial flow compressor testing data. Axial flow compressor rotating stall offline analysis system is user-friendly, easy to operate, providing good assistance and data support for online real-time detection, thereby promoting the process of the project.