Intelligent Quantification of Metal Defects in Storage Tanks Based on Machine Learning

Abstract

Wall-climbing robot are seeing increasing adoption to automated remote and in situ inspection of industrial assets, removing the need for hazardous manned access. The ultrasonic dry-coupling detection device installed on the wall-climbing robot detects the defects of the tank wall. Aiming at the difficulty that the ultrasonic A-scan signal obtained by the ultrasonic dry-coupling detection method has waveform cross-aliasing, which makes it difficult to obtain effective information in traditional feature extraction, Herein, we combine the fast Fourier transform, wavelet packet decomposition and empirical mode decomposition techniques to propose a 3D-SFE method performs multi-scale feature extraction on dry coupled signals. At the same time, in view of the difficulty that traditional nondestructive testing models cannot quantify the defect area accurately, we introduce the XGBoost model to better quantify the defect area. Our proposed defect area quantification model based on multi-scale feature extraction achieves 99.9% accuracy on the training set and 81.5% on the test set. Furthermore, we also analyzed the influence of defect characteristics, sample number, defect shape and depth on the model, and then provided certain guiding significance for the detection of tank defects.