Gaussian process for interpreting pulsed eddy current signals for ferromagnetic pipe profiling

Abstract

This paper describes a Gaussian Process based machine learning technique to estimate the remaining volume of cast iron in ageing water pipes. The method utilizes time domain signals produced by a commercially available pulsed Eddy current sensor. Data produced by the sensor are used to train a Gaussian Process model and perform inference of the remaining metal volume. The Gaussian Process model was learned using sensor data obtained from cast iron calibration plates of various thicknesses. Results produced by the Gaussian Process model were validated against the remaining wall thickness acquired using a high resolution laser scanner after the pipes were sandblasted to remove corrosion. The evaluation shows agreement between model outputs and ground truth. The paper concludes by discussing the implications or results and how the proposed method can potentially advance the current technological setup by facilitating real time pipe profiling.

Designing a pulsed eddy current sensing set-up for cast iron thickness assessment

“…11, i.e., variation of d 2 against β is nonlinear. Such a function may be modeled as a nonlinear regression function using techniques such as Gaussian Process (GP) regression as done in [9], [15], [16], [17]. Despite how a function may be modeled, the prevalence of a functional behavior between cast iron thickness and the β feature is the important and desirable aspect.…”

“…1, the detector coil based PEC sensor is typically composed of two concentrically wound, air cored, conductive circular coils [3], [6], [4]. Concentrically wound rectangular coils too are rarely used [2], [9]. One coil behaves as an exciter coil while the other behaves as a detector coil which captures the signal.…”

Improved signal interpretation for cast iron thickness assessment based on pulsed eddy current sensing

“…In equivalent words, by finding an ALSFL of the logarithmic curve of the measured signal, the thickness of the material under the sensor footprint can be easily obtained by computing a gradient of the ALSFL. Algorithmically, the fitting line from a set of samples can be obtained by the use of Random Sample Consensus (RANSAC) [10], [14]. Nevertheless, this renowned approach is also timeconsuming in computation.…”

“…Generally speaking, due to the nonlinear and heterogeneous ferromagnetic nature of the cast iron, the decaying voltage curves induced by the pick-up coil in the PEC sensor are usually accompanied by noise. In order to denoise the signals, the authors in [5], [10] proposed an average filter that computes an averaged decaying voltage signal from multiple measurements before estimating the thickness of the material. Nevertheless, requiring multiple sensor readings for calculating one thickness in this method is not feasible in realistic applications where testing speed is also a crucial factor.…”

Gaussian Markov Random Fields for fusion in information form

“…We fuse two data sources that contain pipe's remaining wall thickness measurements: 1) a pulsed-Eddy current sensor [23] (abbr. LR sensor) and 2) a magnetic flux leakage sensor [24] (abbr.…”