Analysis Of Acoustic Emission Signal Of Fatigue Crack Based On PCA-DBN

With the rapid development of China's railway industry and the continuous improvement of the highest speed of trains,the safety factor of train axles has become the most important factor in the safe running of trains.Once the axle breaks,it will cause inestimable losses,so it is of great significance for real-time online detection and diagnosis of axle fault.Under the influence of background noise and the impact of axle,how to identify axle cracks quickly and effectively is a great challenge.In this paper,principal component analysis(PCA)and depth confidence network(PCA-DBN)are used to analyze and study various signals under different conditions,and the AE signals of axle fatigue cracks can be identified quickly and effectively.In the experiment of acoustic emission detection,the dimensionless feature parameters and dimensionless feature parameters in the time domain are extracted by sensors to form a fusion feature set,and the trend charts of fusion characteristics of various signals are drawn by MATLAB.At the same time,the PCA method is proposed to obtain the low-dimensional feature sets corresponding to different fatigue states,and map them to a three-dimensional space and two-dimensional space for visualization.The dimension reduction and visualization effects of PCA are better than those of traditional implicit Dirichlet LDA and locally linear embedding LLE methods.Then the PCA-DBN classifier with the optimal structure 3-6-4-3 is used to ensure the effective identification of axle fatigue cracks.In order to break through the viewpoints of randomness and instability of model recognition rate,a new 7 million axle crack data is added to verify the stability of PCA-DBN model(3-6-4-3).The experimental verification and comparison show that the PCA-DBN model(3-6-4-3)structure is very stable for different experimental data and the classification performance of DBN can reach 97.92%.This shows that the classification and recognition effect of PCA-DBN in deep belief network is better than that of PCA-BP neural network(85.42%)and PCA-SVM model in support vector machine(86.46%).It fully proves the necessity and superiority of deep belief network PCA-DBN for classification and recognition.