In order to address the scientific challenge of predicting the remaining useful life of materials damaged by cavitation erosion,a life prediction method that integrates convolutional neural network technology is introduced.Specifically,the residual network(ResNet)model is utilized and enhanced with a coordinate attention mechanism(CA).Through optimization of the convolution,channel number and down-sampling method in the model,an improved coordinate residual network(CA-ResNet)model is developed to accurately predict the remaining useful life of 17-4PH material damaged by cavitation erosion.Initially,the cavitation characteristic curve is obtained from ultrasonic cavitation tests.Then,the cavitation stages are quantitatively segmented using a Logistic equation,defining the life coefficient ζ.Meanwhile,with the assistance of a super-depth-of-field microscope,microscopic images of the material at various post-damage time points are captured to establish a microscopic image database correlated with the life coefficient ζ.The results demonstrate that the improved CA-ResNet network model achieves a verification accuracy of 92.2％on the CIFAR10 public dataset and 93.2％on the collected cavitation damage dataset of 17-4PH material.This represents a 1.5％and 3.5％accuracy improvement over the ResNetl8 network model,respectively.By fine-tuning hyperparameters like learning rate and batch size,the accuracy on the cavitation damage dataset is elevated to 95.0％.In this paper,an end-to-end data-driven approach is adopted to achieve accurate prediction from cavitation damage morphology to post-cavitation life.
NETL
NSTL
万方数据
