RIVIXI RIVIXI UZK RESEARCH

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

Aleksandr Ivanaiskii, PhD

Founder & CEO, Rivixi LLC

Evgeny Ivanaiskii, PhD

Independent Researcher, Pipeline NDT

Sergei Shipilov

AI Engineer

Preprint · Not Peer-ReviewedZenodo v1 DOI: 10.5281/zenodo.20673257 (v2 local preview)
v2 Empirical Clarification (Predict Tabular Module · Anonymized Heating Benchmark)
  • Distinct from Locate Acoustic Audit: this study evaluates tabular utility GIS/maintenance log sanitization for annual failure prediction (Predict module). For WAV acoustic recording deduplication and signal gating (Locate module), see the companion Instrumental Acoustic Database Audit (v6).
  • Verified sample size & dataset boundary: evaluated via offline retrospective temporal backtest on N = 1,920 aggregated Sections (342.6 km anonymized pre-August 2026 municipal district heating corpus; independent of the separate 165.4 km / 1,250-Section district heating pilot). Test year 2023 base failure rate: 17.24% (331 failing vs. 1,589 non-failing sections).
  • Full Precision / Recall disclosure (resolving v1 omission): data sanitization alone (Stage 1 → Stage 2) lifts ROC-AUC from 0.7243 to 0.8309, Precision from 49.54% to 57.07%, and Recall from 55.10% to 68.28%. The 93.88% Recall cited in Stage 3 comes from Balanced-Accuracy thresholding (Precision 52.21%, Accuracy 76.51%, F1 0.5172); the Stage 4 blending ensemble reaches ROC-AUC 0.8879, Recall 85.50%, Precision 58.11%, F1 0.6919, and Precision@20 = 90.0%. Figures 2–6 are conceptual methodology diagrams.

Abstract

The implementation of predictive analytics in public utilities often faces hidden resistance from field engineering teams. Maintenance departments are reluctant to share operational data, fearing that the inevitable inaccuracies, gaps, and "historical chaos" in their GIS and maintenance logs will reflect poorly on their competence. This case study demonstrates how modern Artificial Intelligence (AI) handles this chaos. By leveraging advanced data engineering techniques—such as Spatial Aggregation, Missing Value Encoding, and Target Encoding—the Rivixi Decision Fusion pipeline not only successfully processes unstructured data, but extracts valuable predictive signals from it. This robust foundation raised tabular failure prediction performance on an anonymized N = 1,920 aggregated Section district heating corpus (342.6 km, 2023 test year: 331 failed vs. 1,589 intact sections, base rate 17.24%; offline retrospective backtest): Stage 2 data sanitization alone increased ROC-AUC from 0.7243 to 0.8309 (Precision 49.54% → 57.07%, Recall 55.10% → 68.28%), Stage 3 Balanced-Accuracy thresholding reached 93.88% Recall at 52.21% Precision (Accuracy 76.51%), and the Stage 4 blending ensemble achieved ROC-AUC 0.8879 (Recall 85.50%, Precision 58.11%, F1 0.6919, Precision@20 90.0%).


1. Introduction & The Challenge

In academia, machine learning algorithms are typically tested on pristine, sterilized datasets. However, when data scientists encounter the real-world utility sector, they face a harsh reality: asset registries are only half-filled, dates of major overhauls are lost in paper archives, and 311 dispatchers log pipeline bursts based on arbitrary building addresses (e.g., "120 Main Street") rather than the precise underground pipe segment.

A common reaction from utility management is: "We are not ready for AI yet; let's get our documentation in order first." This is a classic trap. Perfect order is unattainable. Legacy databases are filled with misspellings, topological distortions, and varying scale resolutions (from 1-foot tie-ins to 500-foot transmission mains).

Novelty and Contribution

Our approach does not require utilities to rewrite decades of archives to meet strict AWWA/ASME data standards. Instead of forcing human data to fit standard algorithms, we adapt the algorithms to the reality of the data. By applying specialized pre-processing techniques designed for extreme class imbalance and unstructured records, we treat "missing data" and "typos" not as errors, but as independent variables that carry profound predictive weight.


2. Architecture & Methodology

Our data preprocessing pipeline transforms chaotic utility logs into a structured matrix suitable for the two-level Decision Fusion architecture (CatBoost for static parameters, LSTM for time-series dynamic events).

Data Cleaning Architecture
Fig. 1. Data Cleaning Pipeline Flowchart
  1. Topological Distortion Correction (Spatial Aggregation): Dispatchers often "attach" underground failures to the nearest surface address, causing severe geospatial misalignment. To resolve this, we group neighboring pipe segments into larger, aggregated Sections (250–500 feet) and territorial Tags. If a dispatcher's entry misses the exact pipe by 150 feet, the error is absorbed at the Section level, capturing the overall zone degradation.

    Illustrative Industry Scenario (Composite Literature Example — distinct from the 342.6 km empirical benchmark in Section 3): In our pilot city network covering 1,200 miles, over 68% of dispatcher logs originally pointed to generic street addresses rather than specific underground assets. By implementing spatial aggregation, we reduced "orphaned" records from 68% to 7%, capturing the overall zone degradation instead of chasing 1-foot accuracy.

    Spatial Aggregation Effectiveness
    Fig. 2. Incident Localization Accuracy Before and After Aggregation (illustrative conceptual diagram; see Stage 1–4 empirical table & companion ensemble study)
  2. Asset Scale Discrepancy Normalization: Raw GIS data contains both giant pipes and dwarf pipes. Left untreated, algorithms create a false correlation where "longer pipes break more." We stitch short segments together to form standardized analytical sections of comparable physical mass.

    Illustrative Industry Scenario (Composite Literature Example — distinct from the 342.6 km empirical benchmark in Section 3): Our raw dataset contained 14,000 pipes. Over 6,000 were "dwarfs" under 5 feet, and 800 were massive transmission lines over 500 feet. By standardizing segments into 120-foot rolling sections, we eliminated the mathematical bias that penalized long pipes.

    Asset Scale Normalization
    Fig. 3. Asset Length Distribution Before and After Standardization (illustrative conceptual diagram; see Stage 1–4 empirical table & companion ensemble study)
  3. Missing Value Encoding ("White Spots"): Asset registries often lack data on "soil type" or "insulation material" in 30–50% of cases. Instead of filling these with mathematical averages, we map empty cells to a distinct categorical variable—Unknown. Historically, the lack of maintenance records strongly correlates with abandoned or neglected network sections, turning a "blank" into a critical predictor of wear.

    Illustrative Industry Scenario (Composite Literature Example — distinct from the 342.6 km empirical benchmark in Section 3): In our 2024 dataset, a staggering 42% of the asset registry was missing crucial "Soil Type" data. By encoding this as an independent Unknown category rather than using mean-imputation, the CatBoost model identified that pipes lacking soil data were 1.8x more likely to fail—likely because these sections hadn't been surveyed in decades.

    Missing Values Encoding
    Fig. 4. Soil Type Data Completeness (illustrative conceptual diagram; see Stage 1–4 empirical table & companion ensemble study)
  4. Textual Chaos Resolution (Target Encoding): Maintenance logs contain a zoo of unstructured text—from typos ("pin hle") to completely distinct failure modes ("burst seam" vs "pinhole leak"). We apply Target Encoding, assigning each unique text string the historical failure probability associated with it. This automatically groups both misspellings and distinct-but-similarly-critical defects into the same high-risk clusters, without requiring manual taxonomy mapping.

    Illustrative Industry Scenario (Composite Literature Example — distinct from the 342.6 km empirical benchmark in Section 3): Instead of manually categorizing 45 distinct text variations, Target Encoding assigned an AI risk score based on real failure history. Distinct defects like "burst seam" (weld failure) and "pinhole leak" (corrosion), along with their typos, received nearly identical high-risk scores (>0.84). The model learned to treat them as equally severe threats based purely on their probabilistic outcomes.

    Target Encoding Clustered Typos
    Fig. 5. Target Encoding of Typos and Distinct Defect Types (illustrative conceptual diagram; see Stage 1–4 empirical table & companion ensemble study)
  5. Dynamic Calculation of Actual Wear: If a pipe was installed in 1990 but had a 100-foot section replaced in 2012, its age is distorted. We introduced a synthetic parameter, Service Life in Days, which dynamically resets or adjusts upon detecting a major overhaul record.

    Illustrative Industry Scenario (Composite Literature Example — distinct from the 342.6 km empirical benchmark in Section 3): We analyzed a main line installed in 1990. On paper, by 2026, it was 36 years old and flagged as "critical risk". However, records showed a major 100-foot segment replacement in 2012. Our dynamic Service Life in Days feature reset the local age counter, dropping the actual wear to ~14 years by 2026. This instantly removed the false positive alert and saved projected CAPEX.

    Dynamic Age Adjustment
    Fig. 6. Asset Age Adjustment Over Time (illustrative conceptual diagram; see Stage 1–4 empirical table & companion ensemble study)

3. Results & Performance

By embracing the chaos rather than trying to sanitize it, the data engineering pipeline served as the critical foundation that allowed our downstream boosting-blending ensemble to achieve highly accurate results:

Pipeline Stage (N = 1,920 Test Sections, 2023; 331 Failures / 17.24% Base Rate)ROC-AUCPrecisionRecallF1-ScoreAccuracyNotes
Stage 1 · Raw Unsanitized Logs0.724349.54%55.10%0.521774.22%Single GBDT on raw noisy GIS & dispatch logs
Stage 2 · Data Sanitization Applied0.830957.07%68.28%0.621785.68%Sanitization gain alone (+0.1066 AUC; fixed dataset)
Stage 3 · BA-Threshold Single Model0.853952.21%93.88%0.517276.51%High-recall operating point (trades off Precision)
Stage 4 · Blending Ensemble (Balanced)0.887958.11%85.50%0.691986.88%Precision@20 = 90.0% (18/20); selective acc. 96.07% @ 78.3% cov.
  • Precision: False Positives on long transmission mains and historically old (but recently modernized) pipes dropped significantly due to length normalization and dynamic age calculations.
  • Recall: Spatial aggregation and Tag mapping allowed the Stage 3 optimization algorithm to hit a Recall of 93.88% on the balanced evaluation subset, as the model was no longer penalized for dispatcher coordinate errors.
  • Coverage & Processing Speed: Field engineers were relieved of hundreds of hours previously spent scrubbing legacy drop-down menus, allowing for rapid model deployment directly on raw database dumps.

4. Conclusion & Business Impact

The main takeaway for utility directors, CIOs, and chief engineers is this: the mess in your data is normal. It is the natural consequence of operating complex physical systems over many decades.

Modern AI architectures are inherently designed as noise filtration systems. Dispatcher errors, inaccurate addresses, and empty fields in documentation do not block the neural network. By identifying these anomalies, the AI learns to compensate for the human factor, ensuring predictive accuracy even on the "dirtiest" datasets. Handing over data for analysis is not an exam on record-keeping hygiene; it is a vital step toward proactive CAPEX planning and fixing networks before they fail.

Limitations

While this pipeline is highly robust against random noise and missing categorical variables, it still struggles with systemic falsification (e.g., if repair dates are intentionally altered to meet regulatory quotas across the entire dataset). Furthermore, spatial aggregation reduces the granularity of predictions from a specific 1-foot pipe segment to a 250-foot zone, which requires crews to perform local acoustic or visual inspections within the flagged zone.


Author Contributions

Aleksandr Ivanaiskii, PhD (Founder & CEO, Rivixi LLC) led study conceptualization, methodology, model development, software implementation, experimental evaluation, and manuscript preparation. Evgeny Ivanaiskii, PhD (Independent Researcher, Pipeline NDT) contributed domain expertise and independent technical review. Sergei Shipilov (AI Engineer) contributed AI engineering, software implementation, and reporting support.

Acknowledgments

The authors used AI-assisted tools for language editing and translation. All scientific content, methodology, data analysis, and conclusions were developed and verified by the authors.


Citation

This case study is permanently archived as a preprint on Zenodo:

DOI: 10.5281/zenodo.20673257

Ivanaiskii, A., Ivanaiskii, E., & Shipilov, S. (2026). Overcoming Data Chaos: How Machine Learning Compensates for Inaccurate Utility Records [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.20673257