Convolutional Neural Network Based Fatigue Crack Diagnosis

In order to ensure flight safety,it is of great significance to diagnose fatigue crack of aircraft structures.The active guided wave(GW)based method has been widely studied for its advantages of large width of monitoring ranges and sensitivity to small damage.The GW based method usually diagnoses fatigue cracks by extracting damage features of GW signals.In the real application process,GW damage features have dispersion which are affected by crack propagation path,sensors,sensors paste and many other factors.This dispersion confuses the reflection of signal characteristics on crack and reduce the reliability of fatigue crack diagnosis.With the vigorous development of deep learning(DL),it owns the ability of fusing a large number of features,learning and extracting high-level feature expressions related to classification,so as to realize efficient classification.Due to these advantages,DL provides a practical and effective new research method for the fatigue crack diagnosis of aircraft structures.In this paper,the convolutional neural network(CNN)model and its application in the fatigue crack diagnosis of aircraft structures are studied.The main contents of this paper are as follows:(1)The basic idea of CNN is studied.A one-dimensional(1D)CNN is designed,and the GW-CNN based fatigue crack diagnosis method is proposed for fatigue crack diagnosis.DL methods or algorithms such as data standardization,dropout,L2 regularization are adopted to improve the performance of the designed CNN.(2)Fatigue tests are carried out on the attachment lug which is a typical aircraft structure.The active GW based structural health monitoring method is used to monitor the fatigue cracks on-line.The method of extracting damage features is studied,the characteristics of damage features change with crack propagation are analyzed.A method of combining multi-channel and multi-GW characteristics into a data sample is proposed,which allows a sample fuses multi-crack information from the same crack state.(3)The reliability of the proposed method for different test specimens is verified by the data which are obtained from the fatigue test of attachment lugs,and the invariance of the features extracted by CNN is analyzed.In addition,the advantages of multi-channel and multi-GW features fusing method and the designed 1D CNN are verified by comparative experiments.