Study On Real-Time Fault Detection Algorithm For Liquid-Propellant Rocket Engine Turbopump

Development of high thrust rocket is an important part of Chinese major projectsof manned space flight, and advanced research on fault detection and diagnosis of therocket power system is an urgent need. Turbopump is one of the most importantcomponents of the liquid rocket engine, and it is most prone to major faults. Thus,research on turbopump fault detection and diagnosis technology is important to ensurethe major projects of manned space flight implement according to the schedule, and itcan speed up the developing process of the new type high thrust rocket. In resent years,research on real-time fault detection and diagnosis algorithm based on vibration signalis the hotspot and frontier of the health monitoring research of the high speed rotatingmachinery. For a certain type of turbopump, choosing its historical vibrationacceleration signal as test object, the research of the real-time fault detection algorithmwas done. The main research contents are summarized as follows:Based on comprehensive consideration of accuracy, real-time and betimes, twokinds of real-time fault detection algorithm based on frequency domain features wereproposed:1) Real-time fault detection algorithm based on peak ratio of frequency band.The algorithm divides the signal frequency spectrum into several bands, chooses twobands and computes the peak ratio as fault feature, set fixed threshold to detect thefaults. After that, an adaptive threshold algorithm was proposed to improve the originalalgorithm.2) Real-time fault detection algorithm based on protruding frequencycomponents RMS and SVM (Support Vector Machine). Based on frequency bandsdivision, contribution coefficient and contribution limit was proposed. In everyfrequency band, the algorithm selects all the frequency components whose amplitudebeyond the contribution limit, computes their RMS value. After that the algorithmcombine the RMS values of all frequency bands to construct a multi-dimensional vectoras fault feature, use SVM method to detect the faults.For the turbopump vibration signal, the time domain features was selected throughthe linear correlation analysis, stability analysis and fault sensitivity analysis of multiplestatistical characteristics. The features have weak linear correlation, fine stability and high fault sensitivity. Based on this, three kinds of real-time fault detection algorithmbased on time domain features were proposed:1) Adaptive C-SVM algorithm based onRMS. Classifier real-time updating method was proposed based on original C-SVMalgorithm. For every channel of signal, the algorithm computes the signal RMS valuesin the same detection step, and combines the RMS values to construct amulti-dimensional vector as fault feature, and then constructs original SVM trainingsample set and acquires original classifier. The training sample set and classifier wouldbe real-time updated in the fault detection process.2) Multi-feature based fast SVMalgorithm. The algorithm divides every detection step of signal into some averagesegments, computes the RMS,margin factor and kurtosis of each segment and constructa three dimensional vector as fault feature, and chooses all fault feature vectors ofhistorical signal to construct original SVM training sample set; In original trainingsample set, the algorithm chooses boundary samples to construct a new training sampleset, and gets the support vectors and classifier by training it. A fault judgment strategywas proposed too. At last, Classifier real-time updating method was proposed toimprove the original algorithm.3) Algorithm based on wavelet RMS vectors and SVM.For every detection step of signal, the algorithm gets multi-layer detail signals throughDaubechies wavelet decomposition and reconstruction. Then the algorithm dividesevery layer into some average segments and computes RMS value of each segment,constructs multi-dimensional RMS vector as fault feature by extracting RMS values atthe same position in every layer, and extracts all the fault feature vectors of historicalsignal to construct SVM training sample set and then obtains SVM classifier.Based on above-mentioned theoretical study, the historical signals obtained froman institute of aerospace power were chosen to validate the above-mentioned algorithms.The validation experiments used the historical signals to simulate the real-time faultdetection process during the turbopump hot-firing test. The results showed that all thealgorithms met the demands of accuracy, real-time and betimes.At last, the historical signals were also chosen to validate three kinds of traditionalreal-time fault detection algorithm.