Applying Advanced Machine Learning to Enhance Indication Discrimination and Data Analysis Results in Ultrasound Crack Diagnosis Surveys

Abstract In-line inspection (ILI) ultrasound for crack detection is a well-established inspection method providing pipeline vendors with a high probability of detection, identification, and sizing, and a low rate for unnecessary digs. In recent years, manual data analysis work by human analysts has been increasingly supported by algorithms, especially in terms of anomaly detection and classification. More recently, conventional rule-based algorithms are being replaced by algorithms based on machine learning, which can deliver deeper logical connections from the data itself and provide higher efficiency. However, the industry is still cautious about adding extensive automation to the processes. Although, the advantages for pipeline vendors of a higher degree of automation are clear: results are more robust and consistent, the human impact is reduced and the achievable times for priority notifications and final reports can be reduced. Discriminating non-relevant anomalies allows highly skilled human analysts to focus on complex situations, which leads to more reliable results in complex defect and pipeline geometries. In this paper, we will present a successfully applied anomaly discriminator based on machine learning (ML) to reduce the initial false positive rate from a previous automatic detection algorithm in ultrasonic crack technology. The goal of this paper is to achieve a deeper understanding of new algorithmic possibilities and how they are integrated into ILI analysis processes to successively build up trust and reveal their high potential to accomplish better and faster outcomes.