The Aerostat Capsule Defect Detection Based on Strain Data and Improved MHA Model
For the difficulty of detecting the surface defects of aerostat capsules,a defect detection method based on strain time series data and an improved multi-head attention(MHA)model is proposed.This proposed method performs end-to-end feature extraction and detection of strain time series data after applying acoustic ex-citation to the capsule and then gets the results of capsule defect detections.First,strain time series data is col-lected from strain gauges attached to the surface of the capsule at the same location with and without cracks un-der acoustic excitation.Then,these collected strain time series data are divided into samples according to a cer-tain length,while each sample is divided into a combination of multiple time series vectors and input into the im-proved MHA model to extract the defect features hidden in the time series data.Finally,the model outputs the corresponding defect identification results for each time series sample.The detection results of the proposed method are compared with the other four traditional models on the collected capsular strain data,and the pro-posed method achieves an average detection accuracy of 97.7%better than the other four models,which verifies the effectiveness of this method.
aerostatneural networkfault detectioncapsuleattention mechanismtime series
NETL
NSTL
万方数据
