Our previous studies on microwave NDT mainly focus on the localization of defects using time of flight. The resonant frequency of reflected microwaves is related to the wall thinning size since the wall thinning region works like a resonant cavity. Consequently, this paper tackled the sizing problem using resonant frequency as the input of back propagation neural network.
A signal processing method was proposed to extract the wall thinning-related signals with the clear frequency spectrum for further analysis.
Results show a negative relationship between the resonant frequency and the wall thinning size (depth and length). The simulation results agree well with experimental results so that we can use simulation results to train the neural network. The trained neural network performed well on experimental data even though only simulation data were used for training.
Next is to challenge sizing partial circumferential wall thinning. Come on!
Title: Quantitative Evaluation of Pipe Wall Thinning Defect Sizes Using Microwave NDT（マイクロ波を用いた配管減肉形状の定量的評価）
Authors: Yijun Guo, Guanren Chen, Takuya Katagiri, Noritaka Yusa, Hidetoshi Hashizume
Abstract: This study investigated the applicability of microwave nondestructive testing, which has been proved effective in quickly detecting the defect location in a long pipe, to the size evaluation of wall thinning defects. Artificial wall thinning defects with different sizes (depths and lengths) and edge profiles were introduced to a flanged brass pipe with a total length of 15 m, and reflected microwave signals were measured in experiments. A signal processing method combining windowing and dispersion compensation was proposed to extract the defect-related reflection signals in the frequency domain. Resonant frequencies, at which the amplitude of extracted signals dropped significantly, decreased with the increase of either wall thinning depth or length. In addition, the results demonstrate that wall thinning location and pipe end conditions have little influence on resonant frequencies after signal processing. A back propagation neural network was trained by simulation data, using resonant frequencies as the input, to simultaneously evaluate defect depth and length, and the performance was validated by experiments. Maximum prediction errors of depth and length of wall thinning were 0.06 mm and 0.57 mm, respectively, which indicated the feasibility of proposed method to evaluate the wall thinning defect sizes.