RIVIXI RIVIXI UZK RESEARCH
Evidence Base & Preprints

RIVIXI UZK Research Archive

All verified and v2-remediated preprints, technical audits, synthetic benchmarks, and digital-twin case studies consolidated on uzk.rivixi.ai. All studies carry Preprint (not peer-reviewed) status and Zenodo DOIs.

CASE STUDY · DIGITAL TWIN & HYBRID AI Preprint · Not Peer-Reviewed

False Positive in Acoustic Leak Detection of a Flooded Heating Pipeline: A Case Study in Hybrid AI Diagnostics

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · June 23, 2026

Field case study on a flooded 336 m DN400 (426x7 mm) district heating line: third-party correlator and primary AI classifiers flagged a leak, whereas hybrid PNR (8.64) and phase coherence (0.4%) produced the verified no-excavation verdict. Open stereo WAVs included.

PNR: 8.64 · Coherence: 0.4% · L = 336 m · Open WAVs
Read Full Study →
RESEARCH PAPER · PREDICTIVE ML (V2) Preprint · Not Peer-Reviewed

Improving Pipeline Failure Prediction via Data Sanitization, Hyperparameter Optimization, and Boosting Blending Ensembles

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · June 4, 2026 (v2 scope update)

Offline retrospective benchmark on a municipal district heating corpus (342.6 km; 8,243 Sys segments aggregated into 1,920 Sections; Train 2020-2021, Val 2022, Test 2023 with 331 positive sections = 17.24% base rate; not an operational rollout). Boosting blending achieved ROC-AUC 0.8879 (95% CI [0.865, 0.908]) and Precision@20 90.0%.

Heating 342.6 km (1,920 sec.) · ROC-AUC: 0.8879 · P@20: 90.0% · Offline test
Read Full Study →
RESEARCH PAPER · PREDICTIVE ML Preprint · Not Peer-Reviewed

Application of Hybrid ML Models for Pipeline Failure Prediction

Authors: E. Ivanaiskii, I. Nazarov, P. Chistyakov, A. Ivanaiskii · May 20, 2026

Two-level Sys-to-Section hybrid ML model evaluated on a municipal district heating corpus (8,243 Sys identifiers across 1,920 Sections; 2023 test year: 490 section failure events, 331 structurally confirmed pipe failures in training). Achieved ROC-AUC 0.8539, Balanced Accuracy 0.8222, Recall 93.88% (460/490), Precision@100 81%.

ROC-AUC: 0.8539 · Recall: 93.88% (460/490) · Precision@100: 81%
Read Full Study →
RESEARCH PAPER · SYNTHETIC DATA Preprint · Not Peer-Reviewed

Spectral Shortcut Learning and Generalization Degradation in 2D-CNN Spectrogram Classifiers Trained on Physics-Agnostic Procedural Synthetic Data

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · August 17, 2026

10-seed statistical benchmark on held-out physical validation (N=61: 39 leaks, 22 normal): at 20% procedural synthetic contamination, real leak Recall falls from 97.95+-1.54% to 85.13+-2.99% (p=2.45e-7, Cohen's d=4.27) while missed leaks rise 7.25-fold.

10-seed (N=61) · Recall: 97.95% -> 85.13% at 20% synth · p = 2.45e-7
Read Full Study →
RESEARCH PAPER · HYBRID VETO (V2) Preprint · Not Peer-Reviewed

Optimizing Hybrid Acoustic / Vibroacoustic Leak Detection Algorithms for District Heating Networks under Incomplete Spatial Metadata

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · July 18, 2026 (v2 threshold & overlap update)

V2 preprint (all thresholds restored: L0 = 60 m, CNN veto consensus < 50%). On N=110 ingested recordings (30 leaks, 80 non-leaks; subset of the same 113-file 50 Hz archive used in ZCR pre-filtering), hybrid arbitration on the 3-model ensemble reduced FPs from 43 to 34 (-20.9% rel.; FPR 53.75% -> 42.50% = -11.25 pp abs.) at Recall 63.3% (19/30), p = 0.0039; 2.9% of archive files had distance labels.

N=110 (50 Hz archive subset) · FP 43->34 (-20.9% rel.) · Recall 63.3% · 2.9% L-labels
Read Full Study →
RESEARCH PAPER · DSP VERIFICATION (V2) Preprint · Not Peer-Reviewed

Utilizing Zero-Crossing Rate (ZCR) for Acoustic Leak Detection in Pipelines: From Empirical Models to a Physically Grounded DSP Pipeline

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · June 17, 2026 (v2 threshold & overlap update)

V2 preprint (all thresholds restored: Z >= 8.5, ZCR > 0.06, PNR = 15.0). Evaluated on N=113 50 Hz field files (30 leaks, 83 non-leaks — the same field archive whose 110-file subset is tested in acoustic-metadata) using the standalone correlation channel: FPR drops from 90.36% (75/83) to 77.11% (64/83) (-13.25 pp abs. / -14.67% rel. FP reduction), Recall 100.00% -> 93.33%.

N=113 (50 Hz archive) · Correlator Z>=8.5 + ZCR>0.06 · FPR 90.36%->77.11% (-13.25 pp)
Read Full Study →
RESEARCH PAPER · DSP ADAPTER (V2) Preprint · Not Peer-Reviewed

Adapting an Acoustic / Vibroacoustic Diagnostics AI Platform to Legacy Hardware: Dynamic DSP Pipeline Reengineering

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · June 10, 2026 (v2 photo & scope update)

V2 preprint (third-party SebaLog N-3 photo removed; thresholds restored: < 6.5 warning, >= 8.5 leak detection). Pilot cohort N=60 steel pipe SebaKMT 6,554 Hz recordings (400-1000 mm, 35-128 m) shows Z-score 6.35-6.47 (normal) vs 14.19-16.48 (leak); extreme 1,638 Hz domain-shift audit (N=371) drops 1D-CNN sensitivity to 15%. Validated on 50 Hz steel grids.

Pilot N=60 steel (6,554 Hz) · Z<6.5 vs >=8.5 · 1,638 Hz domain shift: 15% sens.
Read Full Study →
CASE STUDY · WALL THINNING (V2) Preprint · Not Peer-Reviewed

Vibroacoustic Diagnostics of Degradation and Wall Thinning in Pressurized Pipelines: From Physical Wave Propagation Models to a Hybrid DSP-Deep Learning Pipeline

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · June 20, 2026 (v2 single-line scope update)

V2 preprint (threshold p_zscore > 15 restored; scope tightened to single-pipeline DN400 field case study, 426x7.0 mm, L=336 m). Reconstructed wall loss 10.0-18.5% compared against 3 local UT access points (Points 24, 31, 36); max relative deviation |e_est - e_UT| / e_UT <= 3.5% at those reference points (not multi-line blind UT validation).

Single-line case (DN400, 336 m) · 3 UT points · Max rel. dev <= 3.5% (Z > 15)
Read Full Study →
RESEARCH PAPER · LOGGER TRANSFER Preprint · Not Peer-Reviewed

Evaluating Transferability and Validation Protocols of Convolutional Neural Networks for Acoustic Water Leak Detection Using SebaKMT Logger Recordings

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · October 17, 2026 (v2 scope/author sync)

Empirical transfer and site-disjoint benchmark on 180 SebaKMT logger recordings across 6 site groups: random file splitting yields 100% due to session overlap, while strict site-disjoint TFCNN achieves 87.83±3.22% accuracy (within-fold ROC-AUC 94.28±3.87%) and direct cross-instrument transfer specificity drops 34.90% -> 16.50%.

180 WAVs · 6 site groups · Site-disjoint TFCNN 87.83±3.22% · ROC-AUC 94.28%
Read Full Study →
METHODOLOGY · PREDICT DATA SANITIZATION (V2) Preprint · Not Peer-Reviewed

Overcoming Data Chaos: How Machine Learning Compensates for Inaccurate Utility Records

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · June 11, 2026 (v2 empirical update)

Predict-module data engineering case study on an anonymized 342.6 km municipal district heating dataset (N = 1,920 aggregated Section units; 2023 test year: 331 failed / 1,589 intact, base rate 17.24%; offline retrospective backtest). Stage 2 data sanitization alone raises ROC-AUC from 0.7243 to 0.8309; Stage 3 BA-threshold hits 93.88% Recall at 52.21% Precision; Stage 4 ensemble reaches ROC-AUC 0.8879.

Predict · N=1,920 sections (342.6 km) · Stage 1->2 AUC 0.7243->0.8309 · Stage 4 AUC 0.8879
Read Full Study →
TECHNICAL REPORT V6 · LOCATE ACOUSTIC AUDIT Technical Report v6 · Not Peer-Reviewed

Instrumental Audit of an Operational Acoustic Database for Pipeline Leak Detection and AI Training Dataset Curation

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · November 18, 2026 (Version 6.0)

Instrumental audit of 1,580 operational pipeline acoustic WAV files for cross-correlation localization (Locate) and dual-track AI training dataset curation: cryptographic deduplication isolates 45 synthetic test files and 600 copies (935 unique physical signals: 733 mono, 202 stereo); a delay-adapted >=40 Hz coherence and <1% ADC clipping gate admits 80/148 stereo files; 11/35 pipe-record spans pass (31.4%; 95% Wilson CI 18.6-48.0%) with 10/11 (90.9%) range agreement.

Locate · 1,580 WAVs -> 935 unique -> 80 gated stereo · 11/35 CI [18.6–48.0%]
Read Full Study →
RESEARCH PAPER · WAV AUTHENTICITY (V2) Preprint · Not Peer-Reviewed

Authenticity Verification of Acoustic Leak-Survey Recordings from Signal Microstructure

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · October 24, 2026 (v2 scope update)

V2 preprint (not peer-reviewed) on blind WAV screening across 1,588 16-bit PCM recordings (921 field + 667 synthetic across procedural, Physical Leak, and Advanced Physical Leak families). Within-archive 5-fold CV ROC-AUC is 1.000; on held-out Advanced Physical Leak recall drops to 0.61 (0.50 cutoff) and 0.00 (0.75); 2nd-order Butterworth LPF at 0.97 fN drops recall from 0.941 to 0.000; 4/29 (13.8%) external recordings score >=0.50. Validated on 50 Hz grids; 60 Hz US grids require recalibration.

1,588 PCM (921 field + 667 synth) · In-archive CV 1.000 · Adv. Leak 0.61/0.00 · 50 Hz grid
Read Full Study →
RESEARCH PAPER · SYNTHETIC DATA (V2) Preprint · Not Peer-Reviewed

The Illusion of Generalization: Impact of Ultra-Low Synthetic Contamination on Deep Waveform AI Scratch Training in Acoustic Leak Detection

Authors: A. Ivanaiskii, E. Ivanaiskii, S. Shipilov · August 14, 2026 (v2 update)

V2 preprint (not peer-reviewed) showing 1D waveform classifier scratch training under synthetic contamination: 1% heuristic synthetic WAVs inflate internal accuracy to 99.67% while out-of-domain Recall (WDN Hong Kong, N=140 / 70 leaks) drops 70.00% -> 62.86% (49/70 -> 44/70), 5% drops Recall to 58.57% (41/70), and 20% collapses Recall to 11.43% (8/70).

OOD WDN HK (N=140, 70 leaks) · 1%: 70.00%->62.86% · 20%: 11.43% (8/70)
Read Full Study →