Success Stories
AI-Assisted Revision Comparison for Automotive Wiring Harness Drawings

AI-Assisted Revision Comparison for Automotive Wiring Harness Drawings

Manual comparison of wiring harness revisions was slow and error-prone. iCaptur combined two change detectors to rank likely revision zones, catching 84% of the customer’s marked changes and focusing engineering reviews.

The Challenge

Wiring harness drawings are revised often, and each revision sends an engineer hunting for what actually changed. The routine is to lay old and new sheets side by side, mark every difference by hand and write a change summary. It is slow, and one missed change can carry forward unnoticed.

The drawings themselves made automation harder. Each revision was redrawn instead of modified in its original position, leaving content roughly 50 pixels off between versions and a whole block of general notes relocated, all of which a simple overlay reads as change. The sheets also arrived as 200 dpi raster plots, where fragmented line work and squeezed CAD lettering weakened both image comparison and text recognition.

The sample showed the cost. Plain image comparison lit up 137 of the 232 zones on the sheet, although just 28 held real revision updates. OCR was no steadier, matching just 9 of the 23 and 19 part numbers it read from the two revisions. Engineers needed a way to narrow the search without inspecting every part of every sheet.

Solution

The evaluation used the customer’s manually annotated drawing as a benchmark, a sheet carrying 39 highlighted marks across 31 change zones. The proof of concept focused on identifying and prioritizing likely revision areas to support engineering reviews, not on determining or validating the exact changes.

  • Benchmark-Based Evaluation: Image comparison, the two change detectors and OCR text matching were each scored on how many marked zones they found and how many false alarms they raised.
  • Two Detectors, Ranked Results: Used together, the detectors caught 26 of the 31 marked zones. Zones flagged by both were ranked higher, so the most likely changes appear first.
  • Highlighted, Confidence-Ranked PDF: The latest revision was returned with candidate change areas marked and ranked by detector confidence, giving engineers a visual starting point.
  • Supporting Excel Workbook: For every flagged zone, the workbook shows its rank, whether the two detectors agreed, how much content changed and what text appears only in the current revision, so each flag can be checked against data.
  • Limits Documented Openly: A Read-Me tab explains how the benchmark was run, how each method performed, how many false alarms it produced and where the approach falls short, so reviewers can judge its limits for themselves.

Outcome

The dual-detector approach detected 84% of the customer’s marked change zones (26 of 31). Instead of scanning whole drawings without direction, engineers now begin from a ranked shortlist. The benchmark showed this is a more practical way to narrow the search than plain image comparison, which flooded reviewers with false alarms.

Together, the highlighted PDF and workbook give reviewers a visual reference and supporting data for assessing each candidate change. The evaluation also showed where to go next. To identify changes with greater confidence, the next step is to work from native CAD or vector data instead of relying only on raster images and OCR. The benchmark gives that next phase a measurable baseline to improve on.