Research On Fault Diagnosis Of Aeroengine Gearbox Based On Multi-source Data Mining

As a key transmission component of aero engine,the working state of gearbox directly affects the operation status and performance of the entire engine.Also,all types of its failure occupy the majority of mechanical failure in the engine.Once gearbox fails,the whole transmission system would be paralyzed,even causing catastrophic damage to life and property.Therefore,fault diagnosis of key components in gearbox plays a necessary role in ensuring safety and smooth operation of the engine.Aiming at the four key problems of early fault weakness,poor characteristic adaptive extraction ability,redundancy of fault characteristics and poor generalization ability of pattern classification network,a fault diagnosis method based on multi-source data mining for aviation engine gearbox is presented in this paper,mainly as follows:As for the weak amplitude of characteristic frequency in the early fault signal which is hard to extract,a method of vibration signal enhancement based on maximal overlap discrete wavelet packet transform(MODWPT)and adaptive stochastic resonance(SR)is proposed.Based on multiple single-component signals decomposed by MODWPT,a signal enhancement and reconstruction model of a normal variable-scale SR system is constructed.After that,the parameters of the SR system are adaptively optimized through the particle swarm optimization(PSO)algorithm with ISNR indicator as the adaptability function,finally acquiring the feature enhancement version of the original signal.By referring to a set of early fault data from IMS bearing life cycle data,it is verified that the method proposed in this chapter has significant effect on the enhancement of fault signal characteristics in the early period.Besides,the change of fault characteristic frequency along with its multiplication before and after envelope comparison enhancement shows that the noise in each component is suppressed,and the amplitude at fault frequency is highlighted under the potential function in SR.In regard to the issues such as insufficient precision of weak fault feature extraction and poor completeness of feature parameters under strong background noise,a multi-domain fault feature extraction algorithm based on optimized variational mode decomposition(VMD)and empirical mode decomposition(EMD)by the law of entropy is proposed.The algorithm focuses on the entropy index sensitive to cyclic impact force,searching for the optimal decomposition state of modal decomposition algorithms such as VMD and EMD,effectively promoting the separation of noise from fault characteristics,and obtaining the best fault characteristic information,so as to acquire the feature data for the equipment working state as comprehensive as possible.Through simulation and experimental data verification,it is shown that the improved time-frequency domain signal decomposition algorithm can extract the weak fault characteristic information effectively and clearly under strong background noise.In view of the experience selection of features in traditional methods combining with the deep mining of effective features hidden in data,a weighted kernel principal component analysis(WKPCA)signal feature fusion method based on sensitivity indicators is proposed.Based on a fault feature set consisting of indicators from time and frequency domain,singular values extracted through parameter-optimized VMD and envelope-optimized EMD in time-frequency domain,a multi-measure feature filtering model considering correlation,sample distance,and information indicators is constructed.Considering the variation in both the feature itself and the fault information collected by different sensors,the features are weighted and fused by WKPCA,then the kernel width of the feature sample set is assessed hyper-parameter optimization for the optimal separability.Through the verification of multi-source gear fault data,it is shown that the proposed method can significantly enhance the clustering and separability of the original fault characteristic set,and can effectively improve the accuracy of subsequent fault diagnosis model.In view of the poor generalization ability of pattern recognition network under multi-source heterogeneous fault information,this paper mainly compares three different functional networks: Probability Neural Network(PNN),Long Short-Term Memory(LSTM)and Deep Belief Network(DBN),and then the fault recognition efficiency,accuracy and generalization performance of these different functional networks are investigated.Finally,DBN is chosen as the mode recognition network with the best generalization performance,and the fault diagnosis model of aviation engine gearbox is constructed for this.Besides,the highest stability and diagnostic accuracy of DBN under multiple tests are also verified by fault model testing through mixing data sets of gear and bearing.In addition,a set of fault diagnosis system aviation engine gearbox is developed with MATLAB GUI in this paper,which is able to carry out fault diagnosis of key components in aviation engine including gears,bearings,etc.,as well as to realize early fault warning of the components with test points.The system is characterized by its clear,concise interface and its strong engineering application value.