Conventional ILI tools use odometer wheels to determine the location of identified defects. In addition, typically above ground markers (AGMs) are used to confirm and potentially correct for odometer wheel slippage. Free-floating unconventional ILI tools use information from a variety of sensors to accurately locate defects. Accurately identifying joints is a prerequisite for localization and automatic identification is key for a cost effective inspection.

This work focuses on automating the joint identification process with a neural network. We will present deep learning strategies for discrete feature identification and segmentation in time series data, how those strategies are increasing data processing efficiency, current accuracy and limitations, and normalization strategies for data from multiple sensors.