Ultrasonic metal loss technology is a premium inline inspection technology providing high-resolution direct wall thickness measurement. The underlying principle is used to deliver profound insights into a pipeline state in terms of wall thickness variations like corrosion, also channelling corrosion, and lamination features. The high sensitivity of the technology enables the detection of complex features but also is prone to effects due to challenging pipeline complexities. Highly trained analysts follow well-defined processes to distinguish between safety critical features and 'noise' effects, here automated algorithms are challenged.

Using automated feature detection in inline inspection (ILI), the requirements are strict. The rate of non-detected relevant features needs to go against zero (no false negative), as they represent safety risks. Simultaneously, false call rates by the algorithm (false positives) should not be too high to provide a valid input for manual data analysis or a direct report for the customer.

Deep learning technology, an artificial intelligence method, has become a standard tool for complex image recognition in the last few years. Mimicking complex human pattern recognition, it has been successfully applied in many fields like e.g. medical images (X-Rays, magnetic resonance, etc). Therefore, it has a lot of potential in complex ILI analysis like ultrasonic metal loss measurement which relies heavily on human capabilities.

In this paper, we present a new automatic feature detection approach for ultrasonic metal loss technology based on deep learning (DL). We will outline the approach we followed in the case of model selection and training, the evaluation process established to assure process safety, and the system and process integration. In doing so we provide a complete industry example and template for using state-of-the-art machine learning (ML) technology in complex ILI.