Research On Modified Neural Network For Fault Diagnosis And Performance Prediction Of Aeroengine

Aeroengines are the only power resource of civil aircrafts in flight. To efficiently diagnose andmonitor their failures is important for safety, reliability and economical efficiency of civil aircrafts. Inengine fault diagnosis, a typical and complicated mechanical system can be abstracted from the target.Since the structureof the system is complicated, the diagnosis model is seriously nonlinear, and thediagnosis method is varied, the modeling of engine fault diagnosis becomes complex and difficult.Besides, the integrated errors of measurement can also affect the diagnosis result. So far, researchhotspots in the field of aeroengine fault diagnosis have included the following two aspects: theeffectiveness and global performance of diagnosis method, and the practical applicability of diagnosissystem. The former is to solve the problems of aeroengine fault diagnosis model, and to implementintegrated multi-method for fault diagnosis from single ones. The latter is to move the faultprecautionary point forward to a real-time node, and to introduce trouble-saving and preventivemaintenance into fault diagnosis on the basis of traditional route inspection and aviation accidentinvestigation.Based on ACARSand data decoding of aeroengine state recorded by adigital flight-data recorder/quick access recorder(DFDR/QAR), several key problems are studied about aeroengine faultdiagnosis and state monitoring, which can be concluded as follows:(1) The real-time information provided by ACARS cannot meet the requirements of engine faultmodel for on-line training and real-time diagnosing. However, the data recorded by QAR with a highfrequency is more complete. Therefore, in the modeling process based on function algorithm, amethod is proposed to build sample space using both ACARS and QAR data. Then, concerning datafrom these two sources, the classification characteristics of data frame structure are analyzed.Considering the underlying data coding features of aeroengine on airborne data bus, engine parameterdecoding algorithm is presented based on decoding function. And a real-time and general decodingprocess is realized. Its output can provide essential data for the modeling of aeroengine fault diagnosisand state monitoring.(2)The complexity of the engine system determines the diversity of fault diagnosis.As diagnosisdecision is concerned, decision output obtained by integratingseveral methodsis global, comparedwiththat from single diagnosis decision. Multiplex information source and diagnosisknowledgedesigned duringthe engine fault diagnosis process are fused by different strategies,respectively.Formultiplex information source,a data-layer fusion strategy is adopted, and an adaptive weighted fusion estimation algorithm is proposed.According to the characteristics of engineparameters,the weighted factor is iteratively adjusted, thus realizing the integrated output ofparameters.For diagnosis knowledge, a decision-level fusion strategy is used, anda multiple attributedecision fusion methodbased onHWA operator is employed. Hence, the distributed local decisionknowledge is evolved to global decision knowledge.(3)Using artificial neural method to establish engine fault diagnosis model can break through theperformance bottleneck of traditional mathematical and physical modeling in non-stationary,nonlinear, uncertain complex systems. The method has better approximation and generalizationperformance. Aiming at the fault diagnosis problems, an improved artificial neural networkisproposed,and the ant colony algorithm is used to optimize initial weight vectors in the traditionalalgorithm. Therefore, the disadvantages of slow convergence and training oscillationcan be avoided,which are resulted from subjective random weight determination.Furthermore, in the model training,Levenberg-Marquardt algorithm is introduced, whose nonlinear optimal training rules are used toreplace the gradient descent rules of BP algorithm to reduce the probability of cost function trappingin local minimumpoints during training process.So the convergence speed is improved bycontrollingthe complexity of training algorithm.(4) The key of failure prevention for engines lies in performance prediction, which meansextracting typical information from large amount of operation data to analyze engines’ operationalstate and potential tendency. Therefore, pneumatic performance supervision is developed based onengine EGT margin control. The arithmetic principle for EGMT of testing and launching proceduresis analyzed to recognize the cause of EGMT regression. And the improvement suggestion is alsoprovided. Methods centering the theory of intelligent network are applied for developing pneumaticparameter timing function, including rough set to control redundancy, ant colony and LM algorithmsto train network weights. Momentum method and adaptive adjustment of learning rate are combinedtogether to predict engine pneumatic parameter tendency. Therefore, pneumatic parameters areequipped with good learning and generalizing ability. The proposed methods operate well in tendencyprediction for those nonlinear dynamic systems like pneumatic parameters.(5) The implementation of the above theories and methods are realized in an integrateddeveloping environment. The prototype elaborates major functions related to engine failure andperformance supervision, including engine fault diagnosis, pneumatic performance supervision,performance tendency prediction. Practical testing is also launched.