RCP Track - Team Dynamos

From raw ILI runs to aligned, matched, and forecasted integrity decisions.

Comment

About the Project

Inspiration

Pipeline integrity is a data-heavy problem. We were inspired by how much time engineers still spend manually aligning ILI runs, and how small misalignments can mask real corrosion growth.

The hackathon challenge was clear: if we could automate alignment and matching, we could make integrity decisions faster, safer, and more defensible.

What We Learned

We learned that alignment is the foundational step since everything downstream depends on it.

We also saw how noisy real ILI data can be:

  • Missing clock positions
  • Inconsistent feature names
  • Vendor-specific formats

These factors make “simple matching” infeasible. Building an explainable pipeline helped us keep trust in the results.

How We Built It

We designed a modular pipeline:

  1. Normalization

    • Vendor-specific columns are mapped into a standard schema
    • Clock positions are converted to a 0–360° format

  2. Reference Alignment

    • Fixed points (girth welds, valves, fittings, casings, AGMs) are matched across runs
    • Distances are corrected using these references

  3. Matching

    • Anomalies are paired using:
      • Distance
      • Clock position
      • Feature type
      • Dimensional similarity
    • Matching is solved globally using the Hungarian assignment algorithm

  4. Growth Metrics

    • Per-year growth is computed in:
      • Depth
      • Length
      • Width

  5. Exceptions Handling

    • Explicit tracking of:
      • New anomalies
      • Missing anomalies
      • Unmatchable anomalies

  6. Stretch Goals

    • Clustering (DBSCAN)
    • Weakly-supervised ML matching
    • Growth forecasting
    • Segment-level risk ranking

  7. Dashboard

    • A Streamlit UI makes every step auditable and explainable
    • A Gemini copilot provides fast Q&A over outputs

Technical Highlights (Math)

Growth computation

$$ g_d = \frac{d_2 - d_1}{\Delta t}, \quad g_l = \frac{l_2 - l_1}{\Delta t}, \quad g_w = \frac{w_2 - w_1}{\Delta t} $$

Matching similarity score

$$ S = w_d \cdot \frac{|x_2-x_1|}{\tau_x} + w_c \cdot \frac{|c_2-c_1|}{\tau_c} + w_\ell \cdot \frac{|l_2-l_1|}{\tau_\ell} + w_w \cdot \frac{|w_2-w_1|}{\tau_w} $$

Lower (S) indicates a better match.

Risk scoring by segment

$$ \text{Risk} = z(\text{total anomalies}) + z(\text{new anomalies}) + z(\text{mean growth}) $$

Challenges We Faced

Missing values
Many anomalies lacked clock or dimension data. We had to allow distance-only matching while still penalizing missing fields.

Alignment sensitivity
Small reference errors can cascade into mismatches. We built:

  • Linear alignment
  • DTW alignment
  • Hybrid alignment modes

No labeled data
Without expert matches, we used weak supervision to bootstrap ML matching.

What We’re Proud Of

We turned raw ILI data into a complete integrity workflow:

  • Alignment
  • Matching
  • Growth analysis
  • Clustering
  • Forecasting
  • Risk scoring

All while keeping the system auditable, explainable, and practical for real-world use.

Built With

Share this project:

Updates