An elbow wall thinning diagnosis method by highlighting the stationary characteristics of the operating loop is proposed. The accelerations of curved pipe surfaces were measured in a closed test loop operating at a constant pump rpm, combined with curved pipe specimens with artificial wall thinning. The vibration characteristics of wall-thinned elbows were extracted by using a mel-spectrogram in which modal characteristic variation shifting can be expressed. To reduce the deviation of the model’s prediction values, the ensemble mean value of the mel-spectrogram was used to emphasize stationary signals and reduce noise signals. A convolutional neural network (CNN) regression model with residual blocks was proposed and showed improved performance compared to the models without the residual block. The proposed regression model predicted the thinning thickness of the elbow excluded in training dataset.

A pipe wall thinning diagnosis method based on vibration characteristics is proposed. Elbow specimens with artificial pipe wall thinning were fabricated and combined in a loop. By running a pump in the loop, vibration was induced by flow, and the vibrational signals were measured with accelerometers. The effect of pipe wall thinning on the vibrational signals was investigated by analyzing the spectral data of the acceleration signals. The analyzed vibration characteristics were difficult to observe because the change in characteristics was small. A convolutional neural network (CNN) specialized for data recognition was applied to recognize the small change in vibrational signal resulting from the pipe wall thinning. A regression model based on CNN was chosen to learn the tendency of change in the vibrational signals with varying thinning. The data types advantageous for training the regression model were identified. An early stopping technique using the validation data set was adopted to regularize the regression model. The trained regression model was able to predict pipe thinning.