A Feature-Specific Probabilistic Assessment of Pipeline Defect Size From ILI MFL Signal Using Convolutional Neural Network

Abstract

When estimating pipeline burst pressure, one of the prevalent sources of uncertainty that needs to be factored into the calculation is the model error in the estimation of feature depth and length from the in-line inspection tool. Due to modeling technique limitation, as of today many ILI vendors have feature specific error bounds depending on the morphologies of the corrosion, this error can only be reported to operators as an overall error known as the ILI tool tolerance which is usually obtained from samples of excavation data or pull test data. At the most, the error is reported by classes based on corrosion morphologies specified by Pipeline Operators Forum. For example, a commonly reported corrosion depth sizing specification is ±10% of pipe wall thickness at 80% confidence for the General type of corrosion. This can be interpreted as that the error of each reported depth estimations is assumed to fall in a normal distribution with a mean equal to 0 and standard deviation equal to 7.8% of wall thickness. The shape of the distribution, the mean and standard deviation will then be used as constants to factor in the burst pressure calculation. However, these factors are never constant for a sample of defects in reality. In fact, they ought to be variables on an individual feature basis. An example of such an approach would be a feature specific error tolerance, this could be that the estimated depth of a feature is 36%wt in an interval of [30%, 48%] of wall thickness with 80% confidence. This is believed to greatly reduce the level of uncertainty when it comes to failure pressure estimation or other type of pipeline risk assessment. The advancement in Machine Learning today, deep learning with deep neural networks, allows feature-specific error tolerance to be obtained after analyzing visual imagery of MFL signal. In this paper we will describe a novel approach to predict the size of metal loss defects and more importantly the distribution associated with each prediction. We will then discuss the benefits of this approach has with respect to risk assessment such as failure pressure estimation.