pith. sign in

arxiv: 1907.01184 · v1 · pith:REAWC2WFnew · submitted 2019-07-02 · 💻 cs.LG · cs.SY· eess.SY· stat.AP· stat.ML

Gaussian Mixture Marginal Distributions for Modelling Remaining Pipe Wall Thickness of Critical Water Mains in Non-Destructive Evaluation

Linh Nguyen , Jaime Valls Miro , Lei Shi , Teresa Vidal-Calleja This is my paper

Pith reviewed 2026-05-25 11:15 UTC · model grok-4.3

classification 💻 cs.LG cs.SYeess.SYstat.APstat.ML
keywords gaussian mixture modelsremaining wall thicknessgaussian processesnon-destructive evaluationwater mainsmarginal distributionspipe inspectioncondition assessment
0
0 comments X

The pith

Gaussian mixture models fit marginal distributions of remaining pipe wall thickness to support Gaussian process predictions on partial scans.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes using Gaussian processes to infer remaining wall thickness at unscanned sections of water mains after only partial inspection. Because real RWT data does not follow the normal distribution that GPs assume, the authors demonstrate that Gaussian mixture models effectively model the marginal distributions and capture the probability of any RWT value. The method is tested on actual measurements from a cast iron pipeline in Sydney. A reader would care if this enables faster and more practical robotic inspections of critical infrastructure without needing complete dense scans.

Core claim

Gaussian mixture models are proven to fit marginal distributions to the inspected RWT data, effectively capturing the probability of any RWT value and thereby allowing Gaussian processes to infer values at unseen pipe sections despite the non-normal character of the measurements.

What carries the argument

The Gaussian mixture model for marginal distributions of RWT, which represents the data as a sum of Gaussians to match the observed non-normal distribution.

Load-bearing premise

Real RWT measurements do not follow a normal distribution, which is required for standard Gaussian process predictions.

What would settle it

A comparison on the Sydney pipeline data where replacing the mixture with a single Gaussian produces equivalent or better GP predictions on held-out sections would falsify the necessity of the mixture model.

Figures

Figures reproduced from arXiv: 1907.01184 by Jaime Valls Miro, Lei Shi, Linh Nguyen, Teresa Vidal-Calleja.

Figure 1
Figure 1. Figure 1: Water pipe remaining wall thickness interpretations in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Rapid response thickness tool system [8] (a) and its realistic deployment (b). [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: K-S statistic tests on CDF. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: AIC tests on PDF. It is noticed that it is optimized by 5 and 6 components in the GM marginal distribution for the Pipe 1 and Pipe 2 data sets, respectively. TABLE II: AIC RESULTS Gumbel Weibull Gaussian mixture Pipe 1 7844 8053 7281 Pipe 2 9044 9677 7459 B. Gaussian Process In order to model RWT of a CI water main pipe by em￾ploying GP [6], since the sensor measures an average thickness value in an area o… view at source ↗
Figure 5
Figure 5. Figure 5: Remaining wall thickness maps: Collected data, predicted results and ground truth. Pipe 1 (left column) and Pipe 2 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Root mean square errors. [3] F. Munoz, J. VallsMiro, G. Dissanayake, N. Ulapane, and L. Nguyen, “Design of a lock-in amplifier integrated with a coil system for eddy￾current non-destructive inspection,” in Proc. IEEE Conference on Indus￾trial Electronics and Applications, Siem Reap, Cambodia, June 2017, pp. 1948–1953. [4] N. Ulapane, L. Nguyen, J. V. Miro, A. Alempijevic, and G. Dissanayake, “Designing a p… view at source ↗
read the original abstract

Rapidly estimating the remaining wall thickness (RWT) is paramount for the non-destructive condition assessment evaluation of large critical metallic pipelines. A robotic vehicle with embedded magnetism-based sensors has been developed to traverse the inside of a pipeline and conduct inspections at the location of a break. However its sensing speed is constrained by the magnetic principle of operation, thus slowing down the overall operation in seeking dense RWT mapping. To ameliorate this drawback, this work proposes the partial scanning of the pipe and then employing Gaussian Processes (GPs) to infer RWT at the unseen pipe sections. Since GP prediction assumes to have normally distributed input data - which does correspond with real RWT measurements - Gaussian mixture (GM) models are proven in this work as fitting marginal distributions to effectively capture the probability of any RWT value in the inspected data. The effectiveness of the proposed approach is extensively validated from real-world data collected in collaboration with a water utility from a cast iron water main pipeline in Sydney, Australia.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper proposes partial scanning of critical water mains followed by Gaussian process (GP) inference to predict remaining wall thickness (RWT) at unseen locations. It claims that standard GP prediction requires normally distributed input data, which real RWT measurements violate, and therefore introduces Gaussian mixture (GM) models to capture the marginal distributions of the inspected RWT values. The approach is said to be extensively validated on real cast-iron pipeline data collected in Sydney.

Significance. If the GM step were shown to be necessary and to improve predictive accuracy over standard or warped GPs on non-Gaussian RWT data, the work could support more efficient robotic NDE of aging infrastructure. The manuscript provides no quantitative evidence (metrics, baselines, or error bars) that the GM modeling is required or beneficial, and the stated motivation rests on an incorrect premise about GP assumptions.

major comments (1)
  1. [Abstract] Abstract: the sentence 'Since GP prediction assumes to have normally distributed input data - which does correspond with real RWT measurements' is factually incorrect. Standard GP regression places a GP prior on the latent function and a likelihood (which may be non-Gaussian) on the observations; the input locations (axial or circumferential pipe coordinates) carry no distributional assumption. Non-Gaussian RWT can be accommodated directly via non-Gaussian likelihoods or input warping without an explicit marginal GM step. This premise is load-bearing for the claimed necessity of the GM component.
minor comments (1)
  1. [Abstract] Abstract: the claim of 'extensive validation' on real-world Sydney data is not accompanied by any reported metrics, baselines, cross-validation scheme, or uncertainty quantification, making it impossible to assess effectiveness from the abstract alone.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful review and for identifying the imprecise wording in the abstract. We address this point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the sentence 'Since GP prediction assumes to have normally distributed input data - which does correspond with real RWT measurements' is factually incorrect. Standard GP regression places a GP prior on the latent function and a likelihood (which may be non-Gaussian) on the observations; the input locations (axial or circumferential pipe coordinates) carry no distributional assumption. Non-Gaussian RWT can be accommodated directly via non-Gaussian likelihoods or input warping without an explicit marginal GM step. This premise is load-bearing for the claimed necessity of the GM component.

    Authors: We agree that the stated sentence is factually incorrect and that standard GP regression does not impose normality on the input locations. Our intended meaning was that the GP model in this work employs a Gaussian likelihood on the RWT observations, an assumption violated by the empirically non-normal marginal distribution of real RWT measurements. The GM step is introduced to model that marginal explicitly. We will revise the abstract to remove the erroneous claim, clarify the Gaussian-likelihood assumption, and briefly note the motivation for the GM component. We will also add a short discussion in the main text on the choice of explicit marginal modeling versus direct non-Gaussian likelihoods or warping. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper applies standard GP regression after fitting GM models to marginal RWT distributions. The stated motivation (GP requires normal inputs) is an external modeling choice rather than a self-referential derivation. No equations reduce a claimed prediction to a fitted parameter by construction, no self-citation chains support load-bearing uniqueness claims, and no ansatz is smuggled via prior work. The approach is validated on independent real-world data without the central result being equivalent to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on domain assumptions about data distributions and the suitability of GM models for GP inputs. No free parameters, invented entities, or additional axioms are identifiable from the abstract.

axioms (2)
  • domain assumption GP prediction assumes normally distributed input data
    Explicitly stated in the abstract as the motivation for using GM models.
  • domain assumption Real RWT measurements do not follow a normal distribution
    Stated in the abstract as the reason GM models are required.

pith-pipeline@v0.9.0 · 5724 in / 1110 out tokens · 34732 ms · 2026-05-25T11:15:22.676021+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    A live test-bed for the advancement of condition assessment and failure prediction research on critical pipes,

    J. VallsMiro, J. Rajalingam, T. Vidal-calleja, F. Bruijn, T. Wood, D. Vitanage, N. Ulapane, B. Wijerathna, and D. Su, “A live test-bed for the advancement of condition assessment and failure prediction research on critical pipes,” Water Asset Management International, vol. 10(2), pp. 03–08, 2014

    work page 2014

  2. [2]

    Collaborative research on condition assessment and pipe failure prediction for critical water mains,

    D. Vitanage, J. Kodikara, and G. Allen, “Collaborative research on condition assessment and pipe failure prediction for critical water mains,” Water Asset Management International , vol. 10, pp. 15–18, 2014. (a) (b) (c) (d) (e) (f) Fig. 5: Remaining wall thickness maps: Collected data, predicted results and ground truth. Pipe 1 (left column) and Pipe 2 (r...

    work page 2014

  3. [3]

    Design of a lock-in amplifier integrated with a coil system for eddy- current non-destructive inspection,

    F. Munoz, J. VallsMiro, G. Dissanayake, N. Ulapane, and L. Nguyen, “Design of a lock-in amplifier integrated with a coil system for eddy- current non-destructive inspection,” in Proc. IEEE Conference on Indus- trial Electronics and Applications , Siem Reap, Cambodia, June 2017, pp. 1948–1953

    work page 2017

  4. [4]

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

    N. Ulapane, L. Nguyen, J. V . Miro, A. Alempijevic, and G. Dissanayake, “Designing a pulsed eddy current sensing set-up for cast iron thickness assessment,” in Proc. IEEE Conference on Industrial Electronics and Applications, Siem Reap, Cambodia, June 2017, pp. 901–906

    work page 2017

  5. [5]

    C. E. Rasmussen and C. K. I. Williams, Gaussian processes for machine learning. The MIT Press, Cambridge, Massachusetts, London, England, 2006

    work page 2006

  6. [6]

    Spatio-temporal environmental monitoring for smart buildings,

    L. V . Nguyen, G. Hu, and C. J. Spanos, “Spatio-temporal environmental monitoring for smart buildings,” in Proc. IEEE International Conference on Control and Automation , Ohrid, Macedonia, July 2017, pp. 277–282

    work page 2017

  7. [7]

    Spatial sensor selection via Gaussian Markov random fields,

    L. V . Nguyen, S. Kodagoda, and R. Ranasinghe, “Spatial sensor selection via Gaussian Markov random fields,” IEEE Transaction on Systems, Man, and Cybernetics: Systems , vol. 46(9), pp. 1226–1239, 2016

    work page 2016

  8. [8]

    Rapid response non-destructive inspection robot for condition assessment of critical water mains,

    D. Hunt, M. Hussein, C. Stewart, G. Dissanayake, J. VallsMiro, J. Olson, and R. Rossi, “Rapid response non-destructive inspection robot for condition assessment of critical water mains,” in Proc. Australian Conference on Robotics and Automation , Canterbury, New Zealand, December 2018, pp. 1–7

    work page 2018

  9. [9]

    Extreme value statistics for pitting corrosion of old underground cast iron pipes,

    Z. S. Asadi and R. E. Melchers, “Extreme value statistics for pitting corrosion of old underground cast iron pipes,” Reliability Engineering and System Safety , vol. 162, pp. 64–71, 2017

    work page 2017

  10. [10]

    Extreme value analysis (EV A) of inspection data and its uncertainties,

    D. Benstock and F. Cegla, “Extreme value analysis (EV A) of inspection data and its uncertainties,” NDT&E International , vol. 87, pp. 68–77, 2017

    work page 2017

  11. [11]

    A solution to the inverse pulsed eddy current problem enabling 3d profiling,

    N. Ulapane, L. Nguyen, J. V . Miro, and G. Dissanayake, “A solution to the inverse pulsed eddy current problem enabling 3d profiling,” in Proc. IEEE Conference on Industrial Electronics and Applications , Wuhan, China, June 2018, pp. 1267–1272

    work page 2018

  12. [12]

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

    L. Nguyen, N. Ulapane, J. V . Miro, G. Dissanayake, and F. Munoz, “Improved signal interpretation for cast iron thickness assessment based on pulsed eddy current sensing,” in Proc. IEEE Conference on Industrial Electronics and Applications , Siem Reap, Cambodia, June 2017, pp. 2005–2010

    work page 2017

  13. [13]

    Review of pulsed-eddy-current sig- nal feature-extraction methods for conductive ferromagnetic material- thickness quantification,

    N. Ulapane and L. Nguyen, “Review of pulsed-eddy-current sig- nal feature-extraction methods for conductive ferromagnetic material- thickness quantification,” Electronics, vol. 8, pp. 1–13, 2019

    work page 2019

  14. [14]

    Robotic pipeline wall thickness evaluation for dense nondestructive testing inspection,

    J. VallsMiro, N. Ulapane, L. Shi, D. Hunt, and M. Behrens, “Robotic pipeline wall thickness evaluation for dense nondestructive testing inspection,” Journal of Field Robotics , vol. 35, pp. 1293–1310, 2018

    work page 2018

  15. [15]

    Information-driven adaptive sampling strategy for mobile robotic wire- less sensor network,

    L. Nguyen, S. Kodagoda, R. Ranasinghe, and G. Dissanayake, “Information-driven adaptive sampling strategy for mobile robotic wire- less sensor network,” IEEE Transaction on Control Systems Technology , vol. 24(1), pp. 372–379, 2016

    work page 2016