Intelligent Discrimination Of Failure Modes In Thermal Barrier Coatings Based On Acoustic Emission Signals Analysis

Thermal barrier coatings (TBCs) have been widely used in aviation, andaerospace gas turbine engines, due to their excellent thermal insulation, wearresistance and anti-corrosion. However, because of harsh operation condition&complicated structure, TBCs are easy to have crack in coating surface betweenceramic and bond coating, which will cause the coating peel off eventually. One ofthe main issues of nondestructive evaluation of thermal barrier coating failure is todiscriminate the different damage mechanisms from the detected AE signals. In thispaper, acoustic emission (AE) technique is used to real-time monitoring of the failureprocess and further understands the failure mechanism of TBCs under axial loading.A cluster analysis of AE signal is achieved and the resulting clusters are correlated tothe damage mechanisms of the TBCs under investigation. Feature vector of signal forpattern recognition are extracted based on wavelet packet transform. An unsupervisedmethodology based on BP neural network is developed to realize intelligent signalpattern recognition. The main research contents are stated as follows:Firstly, k-means clustering are used for classification of monitoring AE eventsunder tensile and compress test respectively. Key signal parameters are found out forrepresent different damage model. According to the damage characteristics of thermalbarrier coating system, expected AE signal are produced by tensile and compressloading. Several AE signal descriptors are calculated by the acquisition system foreach AE event: the amplitude, energy, duration, rise time and peak frequency. Thesecollected parameters are used as feature vector in the proposed classification method.The results show that peak frequency will be used to find out which damagemechanism corresponds to each cluster. The frequency spectrum range for AE signalsof surface crack, sliding interface crack, opening interface crack and substratedeformation are0.20~0.25,0.27~0.30,0.43~0.5and0.13~0.16MHZ, respectively.Secondly, the wavelet packet transform is applied to analyze AE signals, and theenergy coefficients are extracted as the signal feature vector. The appropriate waveletfunction and the decomposition scale are selected to analyze four typical AE signalsof TBCs. The figures of AE signal time-frequency distribution are obtained based onwavelet packet transform. The ratio of energy of signals in each component wascompared with the total energy of the calculated AE signal energy. The ratios of energy are used as the characteristic information of the AE signal and the input vectorof BP neural network. The results show that there is an obvious difference on energydistribution of AE signal in each component. The ratio of energy of AE signals canrealize the recognition of failure mode well.Thirdly, the intelligent pattern recognition method is created based on a threelayer BP neural network. All class labeled AE signals are divided into three groupsrandomly, so that about60%of the signals are assigned to the training set,20%to thevalidation set and20%to the test set. After trained175epochs by gradient descentwith momentum algorithm, the mean squared of the network is approach to0.01. Itsaccuracy rate is95%. This shows that the net work have high generalization abilityand reasonable design. Thus the neural network has great value in engineeringapplication.