betaFeedback

WorksComputer Science

Evaluating Near-Perfect Accuracy Claims in Machine-Learning-Based Leak Detection for Two-Phase Pipelines: A Condition-Blocked Audit

Abstract

Reliable leak detection in gas-liquid two-phase pipelines is important for process safety, yet machine-learning studies frequently report near-perfect performance using evaluation protocols that may not represent deployment to unseen operating conditions. This study audits a pressure-based leak-detection model using condition-blocked leave-one-condition-out validation, in which the classifier is tested on a flow condition excluded from training. Accuracy decreased from 93.3% under within-condition evaluation to 59.1% across fifteen unseen-condition tests. Ablation analysis showed that most of this difference arose from the data-splitting protocol, with a smaller contribution from encoding operating conditions as model inputs. Under condition-blocked evaluation, the false-alarm and missed-detection rates increased to 34.1% and 10.2%, respectively, demonstrating that near-perfect reported accuracy did not translate into reliable operation at unseen flow conditions. The reduction persisted across all evaluated classifiers, indicating that it was caused by the evaluation design rather than a particular model architecture. Grouped evaluation of an independent acoustic pipeline dataset produced a comparable reduction toward chance-level performance. These findings show that validation units must represent the conditions encountered during deployment. Condition-blocked evaluation and explicit reporting of operational error rates are therefore necessary before machine-learning leak detectors can support process-safety decisions.

Nomenclature

WkW_k Sliding window size (samples)

VSLV_{SL} Liquid superficial velocity (m/s)

VSGV_{SG} Gas superficial velocity (m/s)

KK Number of LOCO cross-validation folds

Δsplit\Delta_{\mathrm{split}} Split-protocol accuracy gap (percentage points)

Δenc\Delta_{\mathrm{enc}} Feature-encoding accuracy gap (percentage points)

II Interaction contrast (percentage points)

FAR False alarm rate

MDR Missed detection rate

FSO Full-scale output

LOCO Leave-One-Condition-Out cross-validation

C0, C1, C2 No-leak, single-point, two-point leakage classes

S1, S2 Upstream and downstream pressure sensors

P1, P2 Feature vectors from S1 and S2 respectively

MC4 Primary cascaded CatBoost model

RMS Root mean square

PSEP Process Safety and Environmental Protection

Introduction

Leak detection in gas–liquid two-phase pipelines is not an academic benchmark problem; it is a process-safety requirement. Undetected leakage events carry consequences ranging from environmental contamination to facility loss, and alarm systems that generate excessive false positives erode operator confidence to the point where genuine events are dismissed. A classifier that reports 100% accuracy during development and then triggers one in three alarms falsely during field operation is not a marginal failure — it is a systemic one. The distinction between within-condition performance, where the model is tested on conditions it has already seen during training, and unseen-condition generalisation, where it is not, is therefore a process-safety question, not merely a machine learning benchmarking question. That distinction is what this paper addresses.

Machine-learning classifiers applied to dynamic pressure signals from two-phase pipelines have recently reached or approached 100% reported classification accuracy [1, 11]. The study subjected to formal evaluation here [1] was conducted within the same research programme as this paper, using the same six-metre horizontal air–water flow loop. Re-evaluating the published claims on the same data and apparatus, under a protocol that requires the model to classify windows from a flow condition it was not trained on, reduces accuracy from 93.3%±0.7%93.3\% \pm0.7\% (within-condition) to 59.1%±0.6%59.1\% \pm0.6\% on the full 15-condition dataset and to 50.9%±0.5%50.9\% \pm0.5\% on the nine conditions matching the original study design. The false alarm rate under this condition-blocked protocol rises from 0.0% to 34.1%.

The pattern is not confined to this apparatus or to leak detection. In ship fuel-consumption modelling using 1 Hz operational data, Chittamuru et al. [4] likewise showed that random train–test splitting admits temporal leakage and yields optimistic performance relative to time-aware validation. In wind-speed forecasting, an inappropriate cross-validation design inflated reported performance by 20.5% compared with a leakage-free protocol [2]. In repeated-measures biomechanical studies, switching to participant-level cross-validation reduced reported performance from 0.91 to 0.66 [3]. In geospatial modelling, random train–test splitting overestimates performance by up to 0.16 AUC and 80% relative MAE relative to spatially blocked evaluation [5]. Bergmeir et al. [13] established the theoretical conditions under which cross-validation is and is not valid for autocorrelated time series, underpinning what practitioners have documented empirically across domains. Feature-engineering guidelines have long cautioned that any transformation that sees outside the training fold will produce optimistic reported performance [6], and a systematic review of ML reporting standards found that fewer than one-third of published studies include evaluation-validity elements [7].

No prior study in the multiphase pipeline leak-detection literature has applied a quantitative decomposition of this kind, nor isolated each source of accuracy inflation rather than treating data leakage as a single phenomenon. This paper makes five contributions.

  1. A leakage decomposition framework that quantifies two accuracy-gap components — the split-protocol component and the operating-

  1. condition feature-encoding component — through a complete 2×22 \times2 four-condition ablation (§ 4.7).

  1. Empirical evidence that the two components are approximately additive: an interaction contrast II, measuring the non-additive remainder after combining the split-protocol and feature-encoding effects, remains within ±3\pm3 pp at all three window sizes, compared with the approximately 34 pp total gap.

  1. Documentation of a previously unreported dual-sensor coverage structure: the downstream sensor (S2) carries explicit metadata labels for nine of fifteen operating conditions; the remaining six are treated as inferred and confirmed non-duplicate by pairwise correlation analysis.

  1. A corrected bidirectional noise robustness protocol in which training and test data receive identical synthetic noise, reflecting the field deployment scenario more accurately than the asymmetric approach in the original study.

  1. External evaluation on the independent GPLA-12 acoustic gas-pipeline dataset [14], showing the pattern holds across a different physical system and sensor modality (§ 5.9).

Section 2 reviews related work. Section 3 describes the dataset, the evaluation protocol, and the corrected methodology. Section 4 reports the corrected results. Section 5 discusses engineering and safety implications. Section 8 concludes.

Related Work

Evaluation Validity in Time-Series and Grouped Experiments

Bergmeir et al. [13] established that cross-validation is valid for time-series prediction only when specific conditions are met — conditions violated when windows constructed from the same recording appear in both training and test sets. Şahin et al. [2] quantified the empirical consequence in wind-speed forecasting: a 20.5% accuracy gap between an appropriate and an inappropriate cross-validation design. In maritime fuel-consumption modelling, Chittamuru et al. [4] applied time-aware validation to 1 Hz operational data from the CCGS Sir Wilfrid Laurier, demonstrating that random splitting of strongly autocorrelated records can overstate deployment-relevant performance and alter model rankings. Feature-engineering practice for time-series ML has documented the same risk for decades [6], but the guideline is not uniformly applied in domain-specific literature, including the pipeline monitoring field.

The Grouping-Unit Problem

Spatial autocorrelation causes random train–test splits to overstate geospatial model performance by up to 0.16 AUC and 80% relative MAE compared with spatially blocked evaluation [5]. In repeated-measures biomechanics, participant-level cross-validation reduced reported model performance from 0.91 to 0.66 relative to random window-level splits [3]. In the present experimental setting, two pressure recordings at the same gas–liquid superficial-velocity pair (VSGV_{SG}, VSLV_{SL}) — taken at different times during the same session — share far more statistical structure than two recordings at different conditions taken seconds apart. The independence unit is the condition, not the window. Leave-one-condition-out cross-validation implements this principle for small factorial experimental designs; it is not proposed here as a novel algorithm.

Reporting Standards

Klement and El Emam [7] reviewed ML reporting guidelines across medical research and found that evaluation-validity elements appear in fewer than one-third of published studies. The present paper addresses this gap by providing a quantifiable decomposition protocol with each component individually estimated and its interaction assessed empirically.

ML for Multiphase Pipeline Leak Detection

Representative contributions to ML-based leak detection in two-phase pipelines, spanning experimental, numerical, and data-driven approaches, are summarised in Table 8. The study audited here [1] uses the same apparatus as related work from the same programme: Ferroudji et al. [11] characterised leak impacts using statistical and wavelet-transform methods; [10] extended that work to multi-leak comparison; [12] provided the flow-modelling foundation; and Barooah et al. [9] first applied time-series ML to the same configuration.

AuthorsMethod / ParametersConfigurationdd / LLLeak
Experimental modelling
Siebenaler et al. (2017)Gamma densitometer, hydrophones, high-speed cameraHoriz. slug, air–water127 mm / 6 m3.2–9.5 mm
Stajanca et al. (2018)Fibre optic, acoustics, accelerometerHoriz. air114 mm / 38 m1–8 mm
Ji et al. (2018)Acoustics, dynamic pressureHoriz. strat./wavy/slug80 mm / 14.3 m3–6 mm
Li et al. (2022)Dynamic pressure, acousticsHoriz. stratified40 mm / 48 m8 mm
Meng et al. (2024)Ultrasonic, pressure, mass flowHoriz. strat./slug/annular32 mm / 4.8 m3 mm
Khan et al. (2024) [8]ERT (void fraction)Horiz. plug/bubbly50.8 mm / 6 m1.8–3 mm
Li and Yao (2021) [14]Acoustic emission; 12 mic./noise combos; external evaluation hereHoriz. gas pipeline—1 pt
Numerical modelling
Adegboye et al. (2021)Pressure, void fraction (CFD)Horiz. stratified60 mm / 18 m1.5–18 mm
Vandrangi et al. (2021)Pressure, temp., volume (OLGA)Horiz. two-phase203 mm / 100 m—
AuthorsMethod / ParametersConfigurationdd / LLLeak
Lin et al. (2024)Pressure, void fractionSubsea jumper48 mm / 3.6 m15 mm
Ren et al. (2024)Void fraction, volume flowHoriz. plug50 mm / 6 m15 mm
Ferroudji et al. [12] [12]Dynamic pressure, hydrophoneHoriz. plug/bubbly50.8 mm / 3 m1.8–3 mm
Machine learning and other approaches
Santoso et al. (2014)ANN (pressure drop)Horiz. plug24 mm / ——
Zhang et al. (2023)MFCC-LSTMHoriz. strat./slug/annular32 mm / 1.2 m2–5 mm
Al-Ammari and Sleiti (2024)ANN (temp., pressure, mass flow)Horiz. multiphase320 mm / 13 km30 mm
Barooah et al. [9] [9]Mechanistic + data-driven hybridHoriz. multiphase76–172 mm / 0.6–6.5 km12–76 mm
Ferroudji et al. [11]† [11]Dyn. pressure; k-NN, SVM, RF, stacking; unspecified CV. ML5: 81.8% reported (our LOCO: 62.6%). ML6: 89.5% reported; not replicable—DP/NF1 sensors absentHoriz. air–water50.8 mm / 6 m1.8–3 mm
Alquennah et al. [1]† [1]Dyn. pressure; cascaded CatBoost; time-based split; reported 100%Horiz. air–water50.8 mm / 6 m1.8–3 mm
This studyDyn. pressure; LOCO; 2×2 ablation; 34 pp gap; ∣I∣≤3|I| \le3 pp; accuracy 50.9–59.1%Same as †50.8 mm / 6 m1.8–3 mm

Table 8. Contributions to leak detection research in multiphase pipeline systems. † marks studies subject to the evaluation-protocol audit in this paper.

Methodology

Experimental Apparatus

The experimental setup is shown in Figure 1. The apparatus used in [1] is a closed-loop horizontal air–water two-phase flow pipeline with internal diameter 50.8 mm and total length 6 m. A leak zone containing two orifice ports (diameters 1.8 mm and 3 mm) is located 3.85 m downstream of the inlet. Three leakage conditions are studied: no leakage (C0), activation of the 1.8 mm orifice only (C1, single-point leakage), and simultaneous activation of both orifices (C2, two-point or multiple leakage, following the terminology of [1]). Flow regimes across the experimental conditions span plug and slug patterns.

Schematic of the experimental two-phase flow loop

Figure 1. Schematic of the experimental two-phase flow loop (50.8 mm ID, 6 m): pressure sensors S1 (2.4 m from inlet) and S2 (5.9 m from inlet), and the two-orifice leak zone (1.8 mm and 3 mm) at 3.85 m. Dynamic pressure sampled at 2 Hz. Only S1 and S2 pressure signals were used in this study.

Dynamic pressure was recorded using two piezoelectric transducers: sensor S1 at 2.4 m upstream of the leak zone and sensor S2 at 5.9 m downstream. Both sampled at 2 Hz. Each experimental condition was held at steady state for 30 s per recording.

Dataset Structure and the Data Hierarchy

The term “sample” refers to different entities at different levels of the data hierarchy, and these must be distinguished precisely.

  • Operating condition: A unique (VSG,VSL)(V_{SG}, V_{SL}) combination. Fifteen conditions were defined—three gas velocities (VSG∈{0.063,0.12,0.19} m/s)(V_{SG} \in\{0.063, 0.12, 0.19\}\ \mathrm{m/s})

and five liquid velocities (VSL∈{1.03,1.34,1.63,1.94,2.21}V_{SL} \in\{1.03, 1.34, 1.63, 1.94, 2.21\} m/s).

  • Recording (trace): One 30-second continuous pressure time series at a given condition and leakage class. There is one trace per (condition, class) cell — 45 traces per sensor in total.

  • Raw observation: One pressure reading at 2 Hz. Each 30-second trace yields 61 raw observations (30 × 2 + 1).

  • Constructed window: A sliding window of WkW_k consecutive raw observations from a single trace. At Wk=10W_k = 10, stride 1, one trace yields 61−10+1=5261 - 10 + 1 = 52 windows. Windows from the same trace are correlated pseudoreplicates, not independent experimental runs.

The count of 1,647 in Alquennah et al. [1] refers to raw pressure observations: 9 operating conditions × 3 leakage classes × 61 raw observations per condition–class recording = 1,647 observations per sensor. That study used the nine conditions carrying explicit S2 metadata labels (VSL∈{1.03,1.63,2.21}V_{SL} \in\{1.03, 1.63, 2.21\} m/s). The full 15-condition dataset analysed here contains 15 × 3 × 61 = 2,745 raw observations across both sensors. All sample-count references in this paper use “raw observations” for the 2-Hz pressure readings and “windows” for the constructed feature instances.

Sensor coverage. The S2 source files carry explicit VSL metadata labels for nine of fifteen conditions. The remaining six pressure-signal series lacked operating-condition labels but contained valid numerical measurements. These unlabelled series are treated as inferred operating conditions based on pairwise Pearson correlation analysis within each VSGV_{SG} group (maximum observed correlation r=0.837r = 0.837, far below the r>0.999r > 0.999 threshold that would indicate duplication) and monotonically increasing column means consistent with the pressure–velocity relationship in two-phase flow. Because the original operating-condition metadata were unavailable, the six inferred assignments could not be independently verified. Results based on all 15 conditions are therefore reported alongside the nine-condition analysis as a sensitivity assessment. Alquennah et al. [1] used the nine explicitly labelled conditions (POOL9); the full 15-condition pool (POOL15) forms the primary analysis here.

Evaluation Protocol Characteristics in the Original Study

Re-examining the original data-processing pipeline identified four protocol characteristics that contribute to the reported accuracy.

  1. Within-condition evaluation. The published protocol uses an 80/20 time-based split that retains all operating conditions in both training and test sets. This measures within-condition performance: can the model classify later recordings from a condition it has already seen? It does not measure transfer to an entirely new condition. The present study compares this estimand with a condition-blocked LOCO estimand designed to measure that transfer.

  1. Operating-condition feature encoding. Ground-truth VSLV_{\mathrm{SL}} and VSGV_{\mathrm{SG}} values were included as direct model inputs. Including these as features is not inherently inappropriate: if the deployed system has access to flow-meter readings or set-point data, they are legitimate covariates. What is tested here is whether performance is maintained when those features are withheld — a relevant question for deployment scenarios where the precise operating condition is unknown. This is properly described as a feature-ablation test.

  1. Asymmetric noise evaluation. Noise robustness was assessed by training on clean data and testing on synthetically contaminated data. We replace this with symmetric contamination of both training and test, reflecting the scenario where historical sensor recordings and live inference inputs share the same noise characteristics.

  1. Sensor-coverage structure. S2 carries explicit VSL labels for nine of fifteen conditions, as described above. The present analysis includes all fifteen.

Corrected Machine Learning Pipeline

The corrected pipeline (Figure 2) and the two-stage cascade architecture (Figure 3) follow the structure of [1] with the protocol modifications enumerated above.

Corrected ML audit pipeline for multiphase leak classification

Figure 2. Corrected ML audit pipeline for multiphase leak classification (C0: no leak; C1: single-point; C2: two-point/multiple leakage). Two protocol changes are enforced relative to [1]: windows are extracted within individual 30-second recordings only, and LOCO withholds each (VSG,VSL)(V_{SG}, V_{SL}) condition from training in turn. StandardScaler is fitted on the training split of each fold only.

Two-stage cascaded classifier architecture: input feature vector with 16 features ($P1 \times 8$, $P2 \times 8$); Stage 1 – binary classifier, CatBoost · C0 (no leak) vs C1/C2 (any leak); no leak leads to Output: C0, No leak, FAR reported here (false alarms); leak detected leads to Stage 2 – count discriminator, CatBoost · C1 single vs C2 two-point/multiple; C1 leads to Output: C1, Single leak, and C2 leads to Output: C2, Two-point/multiple; MDR measured across C1+C2 (missed detections); Outer LOCO loop: above cascade repeated for each of 15 held-out ($V_{\mathrm{SG}}$, $V_{\mathrm{SL}}$) folds, results aggregated across 10 independent seeds · mean ± std reported; Performance metrics: Accuracy · Macro-F1 (C0/C1/C2), FAR · MDR · per-fold condition breakdown

Figure 3. Two-stage cascaded classifier architecture (MC4). Stage 1 performs binary detection (C0 versus C1 or C2); windows classified as leaking are forwarded to Stage 2, which discriminates single-point (C1) from two-point/multiple leakage (C2). FAR is measured at Stage 1; MDR is measured across Stage 2. The LOCO loop repeats both stages for each of the fifteen held-out conditions, aggregating results across ten independently seeded trials.

Feature Extraction

Eight statistical features are extracted per sensor per window: mean, minimum, maximum, standard deviation, RMS, range, linear-fit slope, and mean absolute difference. These match the features used in [1], allowing direct methodological comparison. Feature vectors are standardised using a StandardScaler fitted exclusively on the training windows of each LOCO fold; the same parameters are applied to validation and test windows without refitting, so no statistical information from the held-out condition enters the preprocessing step.

Leave-One-Condition-Out Protocol

In each of the fifteen LOCO folds, all windows associated with one (VSG,VSL)(V_{SG}, V_{SL}) condition are withheld as the test set; the remaining fourteen form the training set. Within the training fold, an 80/20 stratified random partition produces inner training and validation subsets for hyperparameter selection and early stopping. This inner split operates at the window level, which allows adjacent overlapping windows from the same recording to appear in both inner subsets, repeating in miniature the within-condition dependence documented as Issue 1. The hyperparameter sensitivity analysis (§ 5.3) shows a modest practical impact (approximately 2.3 pp at the available sample sizes), but the tuning results should be treated as potentially optimistic. A recording-block or condition-grouped inner split would eliminate this dependence and is recommended for future studies.

Decomposition Protocol

The split-protocol and feature-encoding contributions are isolated through a complete 2×22 \times2 ablation. Each of the four experimental conditions crosses the two protocol dimensions — split type (within-condition versus condition-blocked) and feature set (with versus without VSLV_{SL} and VSGV_{SG}):

  1. Condition A — published protocol. Time-based 80/20 split; VSLV_{SL} and VSGV_{SG} included as inputs. Reproduces the estimand of [1].

  2. Condition B — split corrected. LOCO applied; VSLV_{SL} and VSGV_{SG} retained as inputs.

  3. Condition C — both corrected (primary result). LOCO applied; VSLV_{SL} and VSGV_{SG} excluded.

  4. Condition D — feature set corrected only. Time-based split retained; VSLV_{SL} and VSGV_{SG} excluded. Without this cell the two components are order-dependent and cannot be assessed separately.

The two accuracy-gap components are defined as:

Δsplit=Acc⁡(A)−Acc⁡(B),(1)\Delta_{\mathrm{split}} = \operatorname{Acc}(A) - \operatorname{Acc}(B), \tag*{(1)}
Δenc=Acc⁡(B)−Acc⁡(C).(2)\Delta_{\mathrm{enc}} = \operatorname{Acc}(B) - \operatorname{Acc}(C). \tag*{(2)}

Approximate additivity is assessed via the interaction contrast

I=Acc⁡(A)−Acc⁡(B)−Acc⁡(D)+Acc⁡(C),(3)I = \operatorname{Acc}(A) - \operatorname{Acc}(B) - \operatorname{Acc}(D) + \operatorname{Acc}(C), \tag*{(3)}

which equals zero when the two components combine without interaction. A small ∣I∣|I| relative to Δsplit\Delta_{\mathrm{split}} and Δenc\Delta_{\mathrm{enc}} supports approximate additivity; a large ∣I∣|I| signals a substantial interaction between the two corrections.

Each experimental condition is evaluated with ten independently drawn random seeds. The primary model, MC4 (cascaded CatBoost), uses the fixed configuration iterations = 300, depth = 6, learning_rate = 0.05. Hyperparameter sensitivity is quantified separately against a nested grid search over 18 combinations ({200, 300} ×\times {4, 6, 8} ×\times {0.03, 0.05, 0.10}). Standard deviations across seeds quantify model-training instability — sensitivity to random initialisation within a fixed data partition — rather than uncertainty across independently repeated physical experiments.

Models Evaluated

The primary model is MC4 (Figure 3), introduced in [1]. For comparison, the seven single-stage classifiers from the original study are also evaluated under the corrected protocol: kk-nearest neighbours (k-NN), support vector machine (SVM), decision tree (DT), random forest (RF), single-stage CatBoost, max-voting ensemble, and stacking ensemble. A feed-forward neural network (256-128-64 units, batch normalisation, dropout rate 0.30, Adam optimiser, patience 15) is included as an additional baseline.

Corrected Noise Robustness Protocol

Synthetic Gaussian noise is added to both training and test feature vectors at identical magnitude, expressed as a percentage of FSO. FSO is derived per feature from the training-fold range with a 10% margin; the noise level is treated as a 3σ3\sigma interval, and all FSO and scaling parameters are estimated from the training fold only. Five levels are evaluated (0.0, 0.1, 0.2, 0.5, 1.0% FSO) under LOCO. Symmetric contamination represents one plausible deployment scenario — it models the case where historical sensor recordings and live inference inputs share the same noise floor. Other scenarios are equally plausible and are not evaluated here.

Performance Metrics

Overall accuracy and macro-averaged F1-score are the primary aggregate metrics. Per-class F1-score is reported for each leakage class. FAR is the proportion of C0 windows classified as leaking; MDR is the proportion of C1 or C2 windows classified as no-leak. All metrics are mean ±\pm standard deviation across ten seeds, where the standard deviation reflects model-training instability, not uncertainty from independently repeated physical experiments.

Results

Complete 2×22 \times2 Ablation

Table 1 and Figure 4 summarise accuracy across all four conditions and three window sizes. At Wk=10W_k = 10, the within-condition accuracy of 93.3%±0.7% (Condition A) falls to 59.1%±0.6% under full correction (Condition C, POOL15_{15}) — a 34.2-percentage-point gap. On the nine-condition sensitivity run matching the original study design (POOL9_9), corrected accuracy is 50.9%±0.5%.

ProtocolWk=6W_k=6Wk=8W_k=8Wk=10W_k=10
A: Time-based + VSL/VSGV_{SL}/V_{SG}89.3 ± 1.0%91.8 ± 0.9%93.3 ± 0.7%
B: LOCO + VSL/VSGV_{SL}/V_{SG}62.1 ± 0.7%61.7 ± 0.7%62.6 ± 0.5%
C⋆^\star: LOCO, no VSL/VSGV_{SL}/V_{SG} (15 cond.)55.4 ± 0.7%56.0 ± 0.8%59.1 ± 0.6%
D: Time-based, no VSL/VSGV_{SL}/V_{SG}84.8 ± 1.2%88.8 ± 0.6%88.4 ± 0.9%
C-9 sensitivity (9 conditions)49.9 ± 0.3%49.8 ± 0.8%50.9 ± 0.5%
Δsplit\Delta_{\mathrm{split}} (A→B)−27.2 pp−30.1 pp−30.7 pp
Δenc\Delta_{\mathrm{enc}} (B→C)−6.7 pp−5.7 pp−3.5 pp
Total (A→C)−33.9 pp−35.8 pp−34.2 pp
Δenc\Delta_{\mathrm{enc}} alone (A→D)−4.5 pp−3.0 pp−4.9 pp
Δsplit\Delta_{\mathrm{split}} alone (D→C)−29.5 pp−32.8 pp−29.3 pp
Interaction contrast II−2.2 pp−2.7 pp+1.4 pp

Table 1. Complete 2×22 \times2 ablation: accuracy (mean ± std across 10 seeds). The interaction contrast I=A−B−D+CI = A - B - D + C quantifies approximate additivity; values ∣I∣≤3\lvert I \rvert\le3 pp at all window sizes indicate the two components combine with negligible interaction.

Accuracy across evaluation protocols and window sizes

Figure 4. Accuracy across evaluation protocols at three window sizes (Wk=6W_k = 6, 8, 10). Error bars: ±1\pm1 standard deviation across ten seeds, reflecting model-training instability. The dashed line marks the accuracy reported in [1]. Condition A uses a time-based 80/20 split retaining all conditions in training and test.

The interaction contrast (Eq. (3)) is −2.2 pp at Wk=6W_k = 6, −2.7 pp at Wk=8W_k = 8, and +1.4 pp at Wk=10W_k = 10. All values are small relative to the dominant split-protocol component of 27–31 pp, supporting approximate additivity of the two corrections. The split-protocol component accounts for approximately 90% of the total gap at Wk=10W_k = 10. The feature-encoding component is larger at shorter window sizes (6.7 pp at Wk=6W_k = 6 versus 3.5 pp at Wk=10W_k = 10), consistent with shorter windows encoding less condition information in their statistical summaries and therefore deriving more discriminative gain from the explicit VSL/VSGV_{SL}/V_{SG} features.

Nine-Condition Sensitivity Analysis

The POOL9_9 run uses the nine conditions with explicit S2 labels, matching the scope of [1] exactly. At Wk=10W_k = 10, corrected accuracy is 50.9%±0.5%50.9\% \pm0.5\% — approximately 49 percentage points below the reported 100.0%. The same conditions, the same 1,647 raw observations, the same sensor configuration: the only difference is whether the model encounters the test condition during training.

Hyperparameter Sensitivity

Comparing fixed default hyperparameters against per-fold nested grid search (18 combinations) at Wk=10W_k = 10 gives a 2.3-point accuracy difference, within one standard deviation of either configuration. FAR and MDR are essentially unchanged. The primary result is not sensitive to the hyperparameter selection strategy at the available sample sizes, though the window-level inner split introduces a residual dependence discussed in § 4.6.

Per-Class Results at Wk=10W_k = 10, Condition C

Table 2 reports per-class F1-score averaged across ten seeds. The no-leak class (C0) achieves the highest F1 (70.8%); two-point/multiple leakage (C2) achieves the lowest (50.3%).

ClassF1 (%)Function
C0: No leak70.8±0.870.8 \pm0.8Binary alarm
C1: Single-point56.6±0.856.6 \pm0.8Count discrimination
C2: Two-point/mult.50.3±0.750.3 \pm0.7Count discrimination
Macro F159.2±0.659.2 \pm0.6
Accuracy59.1±0.659.1 \pm0.6
FAR (laboratory)34.1±1.034.1 \pm1.0C0 classified as leak
MDR (laboratory)10.2±0.410.2 \pm0.4C1/C2 classified as C0

Table 2. Per-class metrics: Wk=10W_k = 10, LOCO, no VSL/VSGV_{SL}/V_{SG}, 15 conditions, 10 seeds. FAR and MDR are laboratory estimates from a single rig; they should not be used as deployment readiness criteria without a prospectively designed field study at the intended operating conditions.

Normalised confusion matrix

Figure 5. Normalised confusion matrix at Wk=10W_k = 10, LOCO, no VSL/VSGV_{SL}/V_{SG}, 15 conditions, averaged over ten seeds. Class C2 is two-point or multiple leakage, consistent with [1]. Overall accuracy: 59.1%; FAR: 34.1%; MDR: 10.2%.

⋆\starMC4 is the pre-specified primary model from [1].

All Models Under the Condition-Blocked Protocol

Table 3 and Figure 6 report accuracy for all nine classifiers at Wk=10W_k = 10, LOCO, no VSL/VSGV_{\mathrm{SL}}/V_{\mathrm{SG}}, POOL15_{15}, 10 seeds each.

ModelAccuracy (%)Macro F1FAR (%)MDR (%)
MC4 Cascade CatBoost (primary)⋆^{\star}59.1 ±\pm 0.60.592 ±\pm 0.00634.1 ±\pm 1.010.2 ±\pm 0.4
MS5 Single-stage CatBoost61.6 ±\pm 0.70.617 ±\pm 0.00827.9 ±\pm 0.610.3 ±\pm 0.9
MS2 SVM57.9 ±\pm 0.00.583 ±\pm 0.00029.4 ±\pm 0.010.3 ±\pm 0.0
MS6 Max-vote ensemble57.7 ±\pm 1.00.581 ±\pm 0.00929.0 ±\pm 0.89.8 ±\pm 0.5
MS7 Stacking ensemble56.4 ±\pm 1.20.564 ±\pm 0.01335.1 ±\pm 1.215.8 ±\pm 1.9
MS4 Random forest55.8 ±\pm 0.60.558 ±\pm 0.00633.0 ±\pm 0.814.4 ±\pm 0.4
MS3 Decision tree53.8 ±\pm 1.20.540 ±\pm 0.01240.6 ±\pm 1.015.2 ±\pm 1.1
NN MLP neural network58.4 ±\pm 1.20.583 ±\pm 0.01316.4 ±\pm 1.98.0 ±\pm 1.0
MS1 kk-NN39.4 ±\pm 0.00.398 ±\pm 0.00057.4 ±\pm 0.017.7 ±\pm 0.0

Table 3. All classifiers at Wk=10W_k = 10, LOCO, no VSL/VSGV_{SL}/V_{SG}, POOL15_{15}, 10 seeds. No model exceeds 61.6% accuracy; the finding is not specific to any architecture. MS1 (k-NN) and MS2 (SVM) are deterministic, hence ±\pm0.0% across seeds.

Accuracy of all nine classifiers at $W_k = 10$, LOCO, no $V_{SL}/V_{SG}$, POOL$_{15}$, 10 seeds

Figure 6. Accuracy of all nine classifiers at Wk=10W_k = 10, LOCO, no VSL/VSGV_{SL}/V_{SG}, POOL15_{15}, 10 seeds. The range 39.4–61.6% confirms the within-condition versus unseen-condition accuracy gap is not specific to the MC4 cascade architecture.

Window-Size Ablation

Table 4: Window-size ablation: Condition C, 15 conditions, 10 seeds. Accuracy does not increase monotonically with WkW_k, in contrast to the within-condition results in [1].

WkW_kDurationAccuracyMacro F1FARMDR
63.0 s55.4±0.7%55.4 \pm0.7\%55.3±0.7%55.3 \pm0.7\%37.7±0.5%37.7 \pm0.5\%10.1±0.2%10.1 \pm0.2\%
84.0 s56.0±0.8%56.0 \pm0.8\%56.1±0.7%56.1 \pm0.7\%35.3±1.1%35.3 \pm1.1\%9.5±0.1%9.5 \pm0.1\%
105.0 s59.1±0.6%59.1 \pm0.6\%59.2±0.6%59.2 \pm0.6\%34.1±1.0%34.1 \pm1.0\%10.2±0.4%10.2 \pm0.4\%

Table 4. Window-size ablation: Condition C, 15 conditions, 10 seeds. Accuracy does not increase monotonically with WkW_k, in contrast to the within-condition results in [1].

Larger windows capture more of the slug dynamics within each 30-second recording but reduce the number of windows extractable per recording, shrinking the training set under LOCO. Whether the Wk=10W_k = 10 advantage over Wk=6W_k = 6 represents a stable property of the flow physics or a finite-sample artefact cannot be determined from 15 conditions alone.

Corrected Noise Robustness

Table 5: Noise robustness under symmetric contamination: Wk=10W_k = 10, LOCO, no VSL/VSGV_{\mathrm{SL}}/V_{\mathrm{SG}}, 10 seeds. Accuracy is essentially invariant across the tested range (58.5–59.1%); the performance floor is set by the evaluation protocol, not sensor noise at these FSO levels. The 0% FSO entry matches Condition C in Table 1.

Noise (%FSO)Paper (asymm.)Accuracy (%)FAR (%)MDR (%)
0.0100%59.1±0.659.1 \pm0.634.1±1.034.1 \pm1.010.2±0.410.2 \pm0.4
0.1100%58.5±0.858.5 \pm0.831.9±1.031.9 \pm1.010.6±0.510.6 \pm0.5
0.299%58.9±1.158.9 \pm1.131.1±1.731.1 \pm1.710.5±0.410.5 \pm0.4
0.5N/A59.0±0.759.0 \pm0.730.4±0.930.4 \pm0.910.5±0.310.5 \pm0.3
1.0N/A59.0±0.459.0 \pm0.429.8±0.829.8 \pm0.810.1±0.410.1 \pm0.4

Table 5. Noise robustness under symmetric contamination: Wk=10W_k = 10, LOCO, no VSL/VSGV_{\mathrm{SL}}/V_{\mathrm{SG}}, 10 seeds. Accuracy is essentially invariant across the tested range (58.5–59.1%); the performance floor is set by the evaluation protocol, not sensor noise at these FSO levels. The 0% FSO entry matches Condition C in Table 1.

Noise robustness under the corrected symmetric protocol

Figure 7. Noise robustness under the corrected symmetric protocol: accuracy (left) and FAR/MDR (right) versus noise level. Accuracy remains within 0.6 pp of the noise-free baseline across all five FSO levels, confirming the performance floor is set by the condition-blocked evaluation, not by additive sensor noise. FAR decreases from 34.1% to 29.8% as noise increases; MDR is approximately constant at ∼10.3%\sim10.3\%.

RankFeatureImportance (%)
1P2_max41.19
2P1_min8.12
3P2_range7.81
4P2_meandiff6.38
5P1_max6.20
6P2_std4.91
7P1_rms4.39
8P2_rms4.33
9P2_mean3.71
10P1_mean3.23

Table 6. Top-10 CatBoost feature importances at Wk=10W_k = 10, LOCO, no VSL/VSGV_{SL}/V_{SG}. P2_max (downstream sensor maximum) dominates at 41.2%.

CatBoost feature importances at $W_k = 10$, LOCO, no $V_{SL}/V_{SG}$

Figure 8. CatBoost feature importances at Wk=10W_k = 10, LOCO, no VSL/VSGV_{SL}/V_{SG}. Green: upstream sensor S1 features (P1); red: downstream sensor S2 features (P2). P2_max alone accounts for 41.2% of total importance.

Feature Importance

External Evaluation: GPLA-12

To assess whether the accuracy gap under grouped evaluation is specific to this apparatus or reflects a general pattern, the A/B/C protocol (§ 4.7) was applied to GPLA-12 [14], an independent publicly available acoustic dataset: two microphone positions, three gas pressure levels (0.2, 0.4, 0.5 MPa), and two noise conditions, yielding 684 observations across twelve combinations. The grouping variable is the (microphone, noise) pair; the target is gas pressure level; and the grouping unit is the domain, defined by four microphone–noise combinations.

Table 7 shows that random-split accuracy of 96.8%±1.3%96.8\% \pm1.3\% decreases to 33.7%±1.1%33.7\% \pm1.1\% under leave-one-group-out evaluation—approaching three-class chance level (33.3%), with a decrease of 61.4 percentage points. Note that only four domain groups are available, so group-level uncertainty is substantial; 684 acoustic segments do not represent 684 independent domain-transfer experiments. Accuracy near chance level under grouped evaluation suggests limited transfer to an unseen microphone–noise domain under the evaluated feature representation, consistent with the rig-based finding across a different physical system and sensor modality.

ConditionAccuracy (%)
A: random split, with domain features96.8±1.396.8 \pm1.3
B: LOGO (4 groups), with features35.4±2.535.4 \pm2.5
C: LOGO (4 groups), no features33.7±1.133.7 \pm1.1
Gap (A−B)−61.4 pp
Feature-encoding gap (B−C)−1.7 pp

Table 7. External evaluation on GPLA-12 (ten seeds, four leave-one-group-out folds). Only four domain groups provide cross-domain evidence; group-level uncertainty is substantial.

Discussion

Reading the 2 × 2 Table

The interaction contrast I=A−B−D+CI = A - B - D + C (Eq. (3)) measures whether the two corrections combine linearly. At Wk=10W_k = 10, I=93.3−62.6−88.4+59.1=1.4I = 93.3 - 62.6 - 88.4 + 59.1 = 1.4 pp — small relative to the dominant split-protocol component (30.7 pp) and the feature-encoding component (3.5 pp). At Wk=6W_k = 6 and Wk=8W_k = 8 the values are −2.2 pp and −2.7 pp respectively. All are within the seed-to-seed variability, supporting approximate additivity.

The practical consequence is straightforward: because the two corrections interact only weakly, addressing either one in isolation recovers roughly its share of the accuracy gap. Correcting the split protocol alone (A→B) recovers 30.7 pp; correcting the feature encoding alone (A→D) recovers 4.9 pp. Correcting both recovers 34.2 pp. A future study that fixes only the split protocol will achieve most of the gain; one that fixes only the feature encoding will not. This asymmetry — the split protocol driving roughly 90% of the gap while feature encoding drives roughly 10% — is consistent across all three window sizes.

The feature-encoding component being larger at shorter windows (6.7 pp at Wk=6W_k = 6 versus 3.5 pp at Wk=10W_k = 10) is physically interpretable. Shorter windows extract less information from the slug dynamics within each recording, so the explicit VSL/VSGV_{SL}/V_{SG} coordinates carry relatively more discriminative weight at small window sizes.

Detection Performance at Unseen Conditions

The per-class results in Table 2 reveal a performance asymmetry that survives the protocol corrections. The no-leak class achieves F1=70.8%F1 = 70.8\%, while two-point/multiple leakage achieves 50.3%. Relative to count discrimination, binary detection — telling the operator that a leak is present at all — retains more discriminative capacity at unseen operating conditions. This has a physical basis: binary detection asks whether the pressure dynamics between S1 and S2 have been disturbed, and any leak at any location creates a detectable pressure perturbation. Distinguishing one 1.8 mm orifice from two orifices of different sizes is a harder problem at 2-Hz temporal resolution, where differences may lie in wave arrival timing or amplitude ratios that a 0.5-second sample cannot resolve.

Both capabilities nonetheless fall substantially below the originally reported values. The false alarm rate of 34.1% at unseen conditions means more than one in three no-leak windows would trigger a field alarm; the missed detection rate of 10.2% means roughly one in ten genuine leak events would go undetected. These are laboratory estimates from a single six-metre rig at a specific range of air–water velocities. They cannot be transferred directly to any operational deployment without a prospectively designed field study using conditions drawn independently from the intended operating envelope.

The neural network baseline (Table 3) achieves a false alarm rate of 16.4% — roughly half the cascade value — despite similar overall accuracy (58.4%). This suggests that architecture choice affects operational alarm behaviour even when aggregate accuracy is comparable, and warrants dedicated investigation in follow-on work.

Operating Conditions at Chance Level

Per-fold analysis shows that conditions at VSL=2.21V_{SL} = 2.21 m/s with VSG∈{0.12,0.19}V_{SG} \in\{0.12, 0.19\} m/s return accuracy of exactly 33.3% — three-class chance — across all ten seeds. Under this model and feature representation, the 2-Hz dynamic pressure signal at those conditions did not provide transferable discriminative information about the leak state from the training conditions. Whether this reflects a physical limit of low-frequency pressure sensing at high liquid velocities, a shortage of diversity among the training conditions, or a combination of the two, remains an open question. Addressing it requires higher sampling rates, additional sensors, or a richer set of training conditions at high VSLV_{SL}.

On the Use of VSL and VSG as Model Inputs

Encoding operating conditions directly as features is not inherently a methodological error. If the deployed system receives live flow-meter readings or set-point signals, including them as covariates is appropriate. The feature-ablation test (A→DA \to D, encoding corrected while split unchanged) recovers only 4.9 pp at Wk=10W_k = 10, far less than the split correction alone (30.7 pp). The large accuracy in the original study [1] therefore came primarily from within-condition evaluation, not from the operating-condition features.

External Evaluation in Context

The 61.4-percentage-point decrease observed for GPLA-12 under grouped evaluation is consistent with the accuracy decrease obtained on the six-metre horizontal air–water flow loop. However, four domain groups provide limited statistical power at the group level — the ±2.5\pm2.5 pp variance under the LOGO condition with features reflects genuine group-level uncertainty. The result does not constitute independent statistical confirmation; it does show that a model trained on one recording setup failed at a different recording setup for the same target variable, a pattern seen in both experiments.

Diagnostic Checklist for Future Studies

Six checks are derived directly from the protocol characteristics identified in § 4.3. They apply to any ML study on grouped experimental multiphase-flow data.

  1. Withhold each test condition from training; report results under LOCO or an equivalent condition-blocked protocol.

  2. Construct windowed features independently within each fold; never apply feature extraction to the full dataset before splitting.

  3. Report performance with and without operating-condition features; note whether those features would be available to the deployed system.

  4. Verify that unlabelled data columns are genuine independent recordings before treating them as additional conditions.

  5. Apply noise to both training and test when evaluating robustness under the assumption that noise is present at both collection and inference.

  6. Report results under all ablation conditions, not only the best-performing configuration.

Directions for Future Work

Several directions follow from the present results. The most immediate is expanding the condition space, particularly at VSL=2.21V_{\mathrm{SL}} = 2.21 m/s where every architecture tested returns chance-level accuracy at unseen conditions. Whether higher sampling rates or physics-informed features that capture the spatial pressure gradient between S1 and S2 can improve transfer at those conditions is testable and has direct implications for system design.

The symmetric noise protocol models one plausible field scenario. Correlated process noise — slug-induced pressure transients intrinsic to horizontal two-phase flow — may interact with leak signatures in ways that Gaussian noise cannot capture. Extending the protocol to cover asymmetric and correlated noise scenarios would sharpen the picture.

The window-level inner cross-validation introduces a residual within-recording dependence. A recording-block inner split would eliminate it and would allow a cleaner comparison of hyperparameter strategies without the confound.

Finally, the FAR and MDR reported here are laboratory estimates. A prospectively designed field study with conditions drawn from the intended operating envelope and an agreed alarm threshold is the minimum required before any deployment decision.

Literature Survey of Leak Detection Contributions

Table 8 situates this work within published research on leak detection in multiphase pipeline systems. Studies marked † are the subject of the evaluation audit conducted here.

Data and Code Availability

The analysis code used for data loading, feature extraction, model training, evaluation, and generation of the reported tables and figures is available in a public GitHub repository at https://github.com/issahnazif-afk/100-percent-warning-sign. The exact version used for this study is archived at Zenodo under DOI https://doi.org/10.5281/zenodo.24218226. The final numerical results and figures are reported in this article; the repository contains the scripts required to regenerate them from the source data. The raw sensor recordings are not publicly available at this time. Researchers wishing to request access to the recordings may contact the corresponding author.

Conclusion

Five findings follow from this study.

  1. A pressure-based ML leak-detection model reporting 100% accuracy under within-condition evaluation achieves 50.9%±0.5% on the nine-condition subset matching the original study design and 59.1%±0.6% on the full 15-condition dataset under condition-blocked evaluation. The same conditions, the same raw observations, the same sensors: the only change is whether the model encounters the test condition during training.

  1. The within-condition versus unseen-condition accuracy gap of 34.2 pp at Wk=10W_k = 10 decomposes into a split-protocol component (30.7 pp) and a feature-encoding component (3.5 pp). The interaction contrast at Wk=10W_k = 10 is +1.4 pp (−2.2 pp at Wk=6W_k = 6; −2.7 pp at Wk=8W_k = 8), supporting approximate additivity: each correction contributes its share approximately independently.

  1. Under condition-blocked evaluation at Wk=10W_k = 10, the false alarm rate rises from 0.0% to 34.1%±1.0% and the missed detection rate from 0.1% to 10.2%±0.4%. Binary leak detection achieves F1 = 70.8% while two-point leakage discrimination achieves 50.3%. Both fall substantially below the originally reported values, and the laboratory FAR and MDR should not be treated as deployment-ready performance guarantees without a prospective field study.

  1. Conditions at VSL=2.21V_{SL} = 2.21 m/s with VSG∈{0.12,0.19}V_{SG} \in\{0.12, 0.19\} m/s return chance-level accuracy (33.3%) across all ten seeds under every model tested. Under this feature representation and model configuration, the 2-Hz dynamic pressure signal at those conditions did not provide transferable discriminative information about the leak state. This is a physically interpretable result with direct implications for the design of future sensing strategies at high liquid superficial velocities.

  1. On the independent GPLA-12 acoustic dataset — a different rig, sensor modality, and target variable — random-split accuracy of 96.8% decreases to 33.7%±1.1% under grouped evaluation, a 61-percentage-point decrease approaching chance level. The within-condition versus unseen-condition accuracy gap is not specific to this flow rig.

The corrected accuracy of 59.1%±0.6% is a starting point, not a ceiling. Additional experimental conditions — especially at VSL=2.21V_{SL} = 2.21 m/s — and higher-resolution sensing are the most direct paths to improved cross-condition generalisation. The evaluation protocol and diagnostic checklist provided here are designed to make future progress on that goal reproducible and comparable across studies.

CRediT authorship contribution statement

Issah N. Suleiman: Conceptualization, Methodology, Software, Formal analysis, Investigation, Visualization, Writing – original draft. Sormeh Serpoosh: Methodology, Validation, Writing – review and editing. Nadine Elkholy: Investigation, Data curation, Validation, Writing – review and editing. Hicham Ferroudji: Investigation, Resources, Writing – review and editing. Mohammad Azizur Rahman: Supervision, Funding acquisition, Resources, Project administration, Writing – review and editing. Matthew Hamilton: Conceptualization, Methodology, Supervision, Project administration, Writing – review and editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This publication was made possible by the grant NPRP14S-0321-210080 from the Qatar National Research Fund (a member of the Qatar Foundation). The authors would like to acknowledge QNRF for providing support and guidance, without which this work would not have been possible. The authors also acknowledge support from the Qatar Research, Development and Innovation (QRDI) Council through the Climate Change and Environment Call (CCEC), project CCEC02-0216-250065. This work was also partially funded by the Canada First Research Excellence Fund (CFREF) through the Transforming Climate Action program. Statements made herein are solely the responsibility of the authors.

References

  1. [1]A. N. Alquennah, H. Ferroudji, M. A. Rahman, A. Ghrayeb, H. Abu-Rub, and A. Barooah, “Early Leak Recognition and Classifica- tion in Two-Phase Flow Systems based on Machine Learning Using Dynamic Pressure Signals,” IEEE Open J. Instrum. Meas., 2026, DOI: 10.1109/OJIM.2026.3706192.
  2. [2]G. Şahin, F. Kürker, A. Nur, and E. Akın, “Short-term wind speed forecasting using leakage-free time-series modeling and statistical residual evaluation,” Sustainability, vol. 18, no. 11, p. 5623, 2026.DOI
  3. [3]A. Karbalaie, F. Abtahi, and C. K. Häger, “Participant-aware model validation for repeated-measures data: Comparative cross-validation study,” JMIR AI, vol. 5, p. e87728, 2026.DOI
  4. [4]S. R. Chittamuru, A. Akinturk, A. Kennedy, J. Barnes, and M. Hamilton, “Time-Aware Validation of Machine Learning Fuel Consumption Models: Evidence from 1 Hz Operational Data, CCGS Sir Wilfrid Laurier,” arXiv preprint arXiv:2608.16833, 2026. DOI: 10.48550/arXiv.2608.16833.DOI
  5. [5]K. Le Rest, D. Pinaud, P. Monestiez, J. Chadoeuf, and V. Bretagnolle, “Spatial leave-one-out cross-validation for variable selection in the presence of spatial autocorrelation,” Global Ecol. Biogeogr., vol. 23, no. 7, pp. 811–820, 2014.
  6. [6]A. E. Becker, “Assessing data leakage effects on the performance estimates of machine learning classifiers,” M.Sc. thesis, Inst. Inform., Univ. Fed. Rio Grande do Sul (UFRGS), Porto Alegre, Brazil, 2023.
  7. [7]W. Klement and K. El Emam, “Consolidated reporting guidelines for prognostic and diagnostic machine learning modeling studies: Development and validation,” J. Med. Internet Res., vol. 25, p. e48763, 2023.DOI
  8. [8]U. Khan, W. Pao, K. E. Pilario, N. Sallih, M. Sohail, and H. Azam, “Real-time automatic flow regime classification and mapping for vertical pipes using dynamic pressure signals,” Int. J. Multiphase Flow, vol. 189, p. 105252, 2025.DOI
  9. [9]A. Barooah et al., “Investigation of multiphase flow leak detection in pipeline using time series analysis technique,” in Proc. Int. Conf. Offshore Mech. Arctic Eng., vol. 87868, 2024, p. V008T11A047.DOI
  10. [10]H. Ferroudji et al., “Analysis of water-air multiphase flow through a pipeline with two leaks using numerical and experimental methods,” J. Dispersion Sci. Technol., vol. 47, no. 1, pp. 36–47, 2026.
  11. [11]H. Ferroudji et al., “Comprehensive analysis of leak impacts on liquid-gas multiphase flow using statistical, wavelet transform, and machine learning approaches,” Process Saf. Environ. Prot., vol. 194, pp. 825–843, 2025.DOI
  12. [12]H. Ferroudji et al. “Numerical and experimental modeling of two leaks behaviour for water-air multiphase flow through a pipeline,” in ASTFE Digit. Libr., 2024.DOI
  13. [13]C. Bergmeir, R. J. Hyndman, and B. Koo, “A note on the validity of cross-validation for evaluating autoregressive time series prediction,” Comput. Stat. Data Anal., vol. 120, pp. 70–83, 2018.
  14. [14]J. Li and L. Yao, “GPLA-12: An acoustic signal dataset of gas pipeline leakage,” arXiv preprint arXiv:2106.10277, Jun. 2021.arxiv.org/abs/2106.10277

Paper details

vlnc:W_29cxyvap
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents