Quantitative evaluation of defects by neural network

M. Takadoya, M. Notake, Y. Yabe, T. Ogi, Jan Drewes Achenbach

Research output: Contribution to journalArticlepeer-review

1 Scopus citations


A neural network is applied to a quantitative evaluation of defects. The training data for learning of neural network are prepared from a numerical analysis based on an elastic wave theory. The question posed here is to. evaluate the depth of surface-breaking cracks in a structural component. The numerical analysis is carried out for ten cracks ranging from 0.6 mm to 2.4 mm separated by 0.2 mm in depth. Waveforms in the time and frequency domains are calculated by the use of the boundary element method for these ten types of crack depth. The neural network used here is the three layered network of perception type. Each layer is called input, association, and output layer, respectively. In this type of network, we can use a learning algorithm called back error-propagation and the network deduces the waveform characteristics from the training data and induces the evaluation criteria. After the network training by the use of numerical waveforms, the network is used to evaluate the measured data obtained from the ultrasonic measuring system. A sample prepared for the experiments has three types of crack of 1.05mm, 1.49mm, and 2.19mm in depth. Their waveforms in the time and frequency domains are sent to the input layer of the trained network. The network performance for these experimental data is discussed and the versatility of the present method is pointed out.

Original languageEnglish (US)
Pages (from-to)443-451
Number of pages9
JournalNondestructive Testing and Evaluation
Issue number1-6
StatePublished - Jun 1 1992


  • Ultrasonics
  • neural network quantitative evaluation

ASJC Scopus subject areas

  • Materials Science(all)
  • Mechanics of Materials
  • Mechanical Engineering
  • Physics and Astronomy(all)


Dive into the research topics of 'Quantitative evaluation of defects by neural network'. Together they form a unique fingerprint.

Cite this