An Intelligent System Approach to Real-Time Ultrasonic Flaw Classification in Weldments.

Abstract

Various advanced ultrasonic flaw classification approaches have been proposed for the determination of flaw types in weldments. Among them, ultrasonic pattern recognition approaches have seemed to be most promising, and thus extensive investigations have been carried out in the ultrasonic nondestructive evaluation community. In spite of these extensive endeavors, these approaches have not been widely used in the current industrial field applications due to some critical barriers in cost, time and reliability. To reduce such barriers, here we propose an intelligent system approach realized by the novel combination of following four ingredients such as 1) development of a PC-based real-time ultrasonic testing system, 2) construction of an abundant experimental database of ultrasonic flaw signals in weldments, 3) establishment of an invariant ultrasonic pattern recognition algorithm by use of newly proposed normalized features, and finally 4) implementation of this invariant algorithm in the form of intelligent flaw classification software. In this paper we address in detail the proposed approach including its four ingredients and its capability of making the industrial field application of ultrasonic pattern recognition approaches low cost, rapid and reliable in their performance. The performance of this approach has been investigated with abundant ultrasonic signals captured from various flaws in weldments with variation in the operational variables of the ultrasonic testing. Especially, we have examined the capability of this approach for reducing the effect of the operational variables on the classification performance in terms of the class-conditional probability density functions and correct accept rates. The reasonable and consistent performances obtained in this work demonstrate the high potential of this approach to serve as a convenient and robust tool for many real world flaw classification problems in weldments.