Corrosion Detection and Classification using Acoustic Emission and Machine Learning Based Approach

Abstract

Degradation caused by corrosion in complex engineered systems strongly affects the economic and industrial growth of a country. Failure caused by corrosion in industries are the major cause of breakdown maintenance. Corrosion detection and monitoring techniques can diagnose health of industrial structures and reduce their life-cycle cost. Corrosion detection and monitoring techniques can be classified into two categories namely destructive testing and Nondestructive testing techniques (NDT). Various NDT techniques like Ultrasonic, Acoustic emission, Guided waves, Eddy currents, Radiographic testing and Magnetic flux leakage have been applied by researchers for corrosion monitoring. Acoustic emission being a passive NDT technique has greater potential to be used as corrosion detection and monitoring technique. Acoustic emission during the accelerated corrosion testing is a reliable method for corrosion detection, however, classification of these acoustic emission signals by machine learning techniques is still in its infancy. To overcome this problem, machine learning based classifier approach is proposed that extracts the statistical features of the acquired acoustic emission signals from accelerated corrosion testing of mild steel samples and then use these distinct statistical features as inputs to the classifier to classify corrosion and no-corrosion state and further corrosion severity level prediction. Proposed method automatically extracts distinct statistical features like AE Mean, AE Energy, AE RMS, Skewness and Kurtosis of each acquired acoustic signal and then present these distinct features as inputs to classifier for classification. Acoustic emission signals for accelerated corrosion process were acquired using acoustic sensor, NI Elvis kit and LabView interface. Three different algorithms, back propagation neural network, radial basis function neural network and naive bayes classifier have been used as supervised learning algorithms for classification of ‘corrosion’ and ‘no corrosion’ state and corrosion severity level prediction. For multi-class problem, five corrosion severity levels have been made based on the mass loss occurred during accelerated corrosion testing. For bi-classification problem, Naive Bayes, BP-NN and RBF-NN showed accuracy of 98.68%, 98.57%, and 100% respectively. For five-class corrosion severity level problem, Naive Bayes, BP-NN and RBF-NN showed accuracy of 90.4%, 94.57%, and 100% respectively. Radial basis function neural network outperformed the other two classifiers and showed the best classification accuracy for corrosion severity level prediction due to presence of gaussian activation function in network hidden layer neurons.