Ultrasonic signal processing methods for detection of defects in concrete pipes Shivprakash Iyer a , Sunil K. Sinha a, ⁎, Bernhard R. Tittmann b , Michael K. Pedrick b a Department of Civil & Environmental Engineering, The Pennsylvania State University, University Park, PA, 16802, USA b Engineering Nano Characterization Center (ENCC), The Pennsylvania State University, University Park, PA, 16802, USA abstract article info Article history: Accepted 19 June 2011 Available online 10 August 2011 Keywords: Buried wastewater concrete pipe Automated defect detection Discrete wavelet transform Unsupervised clustering Multilayer perceptron A-scan signals Automated inspection systems are important for maintenance and rehabilitation of pipeline systems in North America given their budgetary constraints, demand on providing quality service, and the need for preserving their pipeline infrastructure. Automated ultrasonic signal classification systems are finding increasing use in many applications for the recognition of large volumes of inspection signals. This paper presents an automated signal classification system to process A-scan signals acquired with the Ultrasound transducer from a pipe region of interest (ROI). The overall approach consists of three major steps, preprocessing of the signal, multi-resolution analysis for feature extraction, and neural network classification. Finally, a post processing scheme to interpret the classifier outputs and classify the ROI into an appropriate defect class taking into consideration some a priori knowledge of the problem is developed. The proposed post processing scheme is composed of several steps that combine the statistics from the classification matrix as well as a two- step procedure based on k-nearest neighbor and non-linear regression. The feature extraction, classification and post processing schemes proposed in this paper provide a working proof-of-concept for developing this inspection system into an automated field applicable tool. © 2011 Elsevier B.V. All rights reserved. 1. Introduction Automated pipeline condition assessment is vital to developing a cost effective, efficient and sustainable asset management system. A large number of new technologies such as Sewer Scanner and Evaluation Technology (SSET), laser based scanning systems, etc. have made it possible to obtain high quality images of the interior of buried pipes [1]. In spite of buried imaging technologies making giant strides in recent years, the basic means of analysis are still human-dependent: a qualified technician is required to identify defects on a television monitor. Additionally, a defect that appears on the surface to be insignificant (less than 5 mm mouth opening) might actually exist throughout the thickness of the pipe and will go unrecognized by the current inspection methods. Any crack (or defect) classification system that primarily depends on surface characteristics is incomplete and needs to be complemented with additional depth perception to provide for a reliable, accurate and effective buried pipeline asset management system. The previous paper by the same authors presented the proof-of- concept for an automated buried pipeline condition assessment system that can provide additional depth perception of defects in addition to surface assessments [2]. An ultrasound inspection method was proposed and experimental results were presented that showed that this methodology was successful in generating rich data from representative concrete pipe samples. The automated inspection system was proposed as a two-step procedure consisting of reconnaissance and characterization modes. The rich data generated in the reconnaissance mode needs to be processed intelligently in order to be able to exploit the characterization mode of inspection. Studies have shown that manual ultrasonic inspection can be accurate but is highly variable, depending on the inspection skills, training and emotional status or fatigue of inspectors [3]. Automated signal classification is becoming increasingly important in many commercial applications, including nondestructive evaluation of civil infrastructures. Motivation for the use of such systems arises from the need for accurate interpretation of large volumes of inspection data, and minimizing errors due to human fatigue and subjectivity. Automated signal classification systems have the advantage of detecting flaws and interpreting ultrasonic signals consistently and accurately without human intervention. This paper builds on the previous one by proposing an automated methodology to process ultrasonic signals generated in the reconnais- sance mode and classify them into various defect classes. Section 2 gives a brief introduction to machine learning and reviews work related to its application in the automated defect analysis of civil infrastructures. Section 3 discusses feature extraction and describes a discrete wavelet transform (DWT) based method to extract discriminatory features from ultrasonic signals to aid in defect classification. The neural network algorithms used for classifying ultrasonic signals whose features have been extracted are presented in Section 4. Experimental results are obtained and presented in Section 5. Section 6 introduces a post processing algorithm necessary to trigger the characterization mode of Automation in Construction 22 (2012) 135–148 ⁎ Corresponding author. Tel.: + 1 540 231 9420. E-mail address: ssinha@vt.edu (S.K. Sinha). 0926-5805/$ – see front matter © 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.autcon.2011.06.012 Contents lists available at ScienceDirect Automation in Construction journal homepage: www.elsevier.com/locate/autcon