Introduction
Welding is the only pipeline fabrication process which occurs in the field. Due to the lack of consistency, welds are the most common locations for failures in a pipeline. Even a perfectly executed welding process introduces material heterogeneity, particularly in the heat-affected zone (HAZ) adjacent to the weld. This zone experiences significant thermal and mechanical changes, including grain coarsening, phase transformations, and residual stresses. The HAZ is also particularly susceptible to stress-induced cracking and corrosion.
Additionally, weld defects such as porosity, incomplete fusion, and slag inclusions act as stress concentrators, further increasing the risk of failure at these locations in the pipeline. Detecting these defects is a key objective of pipeline integrity management. While traditional magnetic flux leakage (MFL) devices have made significant strides in the area of weld inspection, they face limitations in handling pipelines with complex geometries.
Remnant magnetometry offers an alternative approach by analyzing the residual magnetic signatures left by welding and other stress-inducing processes. These signatures provide valuable insights into the structural integrity of pipeline welds, enabling the identification of potential defects.
Weld identification
As weld identification lacks precise articulatable attributes which might lend it to imperative programming, a neural network is the most promising solution.
INGU Weld Net – the INGU Weld identification neural network – is an implementation of U-Net, a widely used deep learning architecture that was first introduced in the “U-Net: Convolutional Networks for Biomedical Image Segmentation” paper . We used a training set of 110,000 samples to train the INGU Weld Net which resulted in 95% automated weld recognition. A task that would typically take days is now performed in less than 1 hour.
The results of this work were first presented at the 2023 Pipeline Technology Conference in Berlin, see our paper Pipeline joint identification using neural networks.
Weld classification
The magnetic signatures of welds in a given pipeline will have some generic weld properties and some pipeline specific properties. Using principal component analysis (PCA) – a statistical method that transforms high-dimensional data into a lower-dimensional space while preserving as much variance as possible –welds in a pipeline can be classified and anomalous welds in the pipeline can be identified.
The essence of the method is that several core shapes (eigenwelds) in a weld are identified and for every weld in a pipeline it is determined which linear combination of these eigenwelds describes the weld best. Consequently, welds that cannot be described with the most dominant eigenwelds are anomalous and worth investigating.
More details about this method can be found in our paper Anomalous Weld Identification by Applying Principal Component Analysis to Magnetic Flux Density Data Captured by a Free-Floating ILI Tool.
Client case study
Pipers® inspection results
Pipers® Magnetic Flux Density data.
Cut out of a weld. Left side: inside of the pipeline. Right side: outside of the pipeline.
