Identification Of Abnormal State Of Aeroengine Based On Acoustic Emission

The abnormal states of aeroengines would affect flight safety of aircrafts and cause serious accidents.In this paper,an in-depth study for identification of abnormal states was conducted based on acoustic emission(AE)technologies to provide an important basis for health management and fault diagnosis of aeroengines.The acoustic emission signals under abnormal conditions were collected through acoustic sensors and used to identify inlet distortions and foreign intakes of an aeroengine.In addition,metal deformation damage and crack initiation were identified based on specimens made of TC11 titanium alloy.The main research results are as follows:(1)Research on identification of aeroengine's inlet distortion based on AE signalsThe relationship of the amplitudes of AE signals,RMS voltage(RMS),and average signal level(ASL)with the range of the engine's inlet distortion exhibit that the AE is able to detect early changes in inlet distortion,thereby providing features for identifying engine's abnormal states.In order to further realize early warnings of engine's distortion,the empirical wavelet transform(EWT)was used to extract the features of the AE signals.Considering the over-decomposed signals occurred in the EWT,a new parameter of AE,namely unit parameter entropy,was proposed to recognize the distortions based on Shannon's entropy.After the AE signals were processed by the EWT,the unit parameter entropy of each modal component was calculated separately.Then,those modal components with large entropy values were retained for signal reconstruction,and those with small entropy ones were removed.It can be seen from the reconstructed signal that the frequency moves to higher values as the distortion increases,which can be used to better distinguish the distortion of the engine.Based on the frequency spectrum of abnormal signals obtained from the EWT and Variation Modal Decomposition(VMD),the deep convolutional neural network(DCNN)of Inception V3 was adopted as the model of transfer learning to construct the method for identifying the pattern of engine's abnormal state.The network structure and weight values before the bottleneck layer of the Inception V3 model were kept,and the newly added connected layer and output layer were trained.200 groups of the AE signals with five distortion of 0%,30%,40%,45%,and 50%were selected to complete the model training,respectively.Finally,the accuracy of the model was tested.A total of 10,000 iterations were conducted on the model,and the final identification accuracy is stable at about 97%.(2)Research on time location and identification of different foreign intakes by an aeroengine based on AE signalsFor the first time,experiments based on a real aeroengine in China was performed to identify different foreign intakes,including small gravels,medium gravels,ice cubes,fuses,rubber rings and screws.In order to obtain the accurate features and time-frequency information of the AE signals after the foreign matters are put in,short-time Fourier transforming was performed to calculate the maximum frequency and the total energy of each AE event.The results showed that the time points corresponding to the maximum frequency and the extreme points of the total energy are the moments for the AE signals when the foreign matters were thrown in.Under certain conditions,the frequency spectrum of the total energy can help to more accurately determine the position of the abnormal event.In order to identify foreign matters,the VMD was used to extract the features of the AE signals corresponding to seven types of foreign matters.It was clearly found that when different foreign matters were put in,the frequency distribution of the AE signals and the signal amplitude distribution are different from each other.In order to further accurately distinguish different foreign matters,the multi-scale permutation entropies of the original signals were combined with those of the high-frequency and low-frequency time-domain signals,by which seven different foreign matters were successfully distinguished.(3)Research on identification of deformation damage stages and crack initiation of TC11 titanium alloy based on AE signalsWhen increasing the load,a large number of AE signals would be generated from the elastic-yield stage to the fracture stage of the metal.The AE signals is related to the stage of the metal damage.In the elastic-yield stage,a large number of AE signals with large amplitude were produced,whereas the number and amplitudes of the AE signals were reduced in the strengthening.In the necking,there was only a small number of AE signals.Until in the end of the necking stage,there was a short period when no AE signal would generate,which marks the initiation of cracking and the arrival of the fracture.In the fracture stage,the number of AE signals suddenly increases,and the features of the AE signals were similar to those in the elastic-yield one.The above phenomena reveals that the AE signals have different feature information in different stages.In the paper,a feature parameter of the AE signals,namely the weighted peak frequency energy ratio(WPFE)was defined to identify deformation damage stages and crack initiation of TC11.By calculating the WPFE ratio between the high-frequency and low-frequency components in the signals,the four stages of metal deformation can be distinguished into two categories:one is the elastic-yield stage and fracture stage,and the other is the strengthening and necking stage.By combining the energy value and the peak frequency with maximum energy of the components,the four stages were well distinguishable.Crack initiation occurs in the fracture stag.The accurate identification of the fracture stage can be applied to recognize crack initiation.Transfer learning was applied to identify various stages of deformation and damage and crack initiation of TC 11 titanium alloy in order to artificially monitor the abnormal states and fault diagnosis of aeroengines.The feature vector obtained by transfer learning was combined with the energy value and peak frequency of the signal having maximum energy as the input of Support Vector Machine to identify the four stages of metal deformation and the initiation of cracks.The results showed that the identification accuracy of the deformation damage and crack initiation of TC11 titanium alloy amounts to 96%.