P&ID Line Register Automation

Overview

Objective and scope

The workflow reads a clean P&ID during inference, then preserves raw alternatives, evidence crops, prediction locks, and post-lock scoring. It is intentionally bounded to line-label localization and transcription; it does not infer piping connectivity or deliver a final production line register.

Engineering question

Manual line-label extraction is repetitive, but a seemingly plausible OCR reading can be wrong or out of drawing scope. The pipeline therefore keeps detection, parsing, review, and evaluation traceable rather than hiding uncertainty behind automatic acceptance.

Objectives

• Locate and transcribe candidate line identifiers from a clean P&ID without pre-reading held-out truth.
• Retain evidence for candidate selection, alternatives, crops, confidence flags, and post-lock evaluation.
• Keep a human review gate in front of any register use.

Method

How the work was approached

• Run dual-orientation OCR with high-resolution local refinement and restore candidate geometry to drawing coordinates.
• Apply loose detection, strict parsing, candidate fusion, and review-status flags without accepting records automatically.
• Lock held-out predictions before marked-PDF, Excel, or adjudicated-truth access, then score localization and transcription separately.

Tools

Assumptions and boundaries

• The benchmark is for line-label localization and transcription only.
• Confirmed visible-label extents are evaluated separately from eight ambiguous occurrences that need more adjudication.
• AUTO is disabled; auto-eligible is diagnostic only and accepts no production record.

Evidence

What the public case is based on

• V1 README, artifact inventory, and pipeline modules for OCR, parsing, fusion, geometry, inference, and validation.
• Held-out benchmark protocol and validation report.
• Evaluator V1.2 audit and comparison, including its further-adjudication decision.

Validation

• The validation report confirms held-out prediction locks preceded first ground-truth access and all 25 unit/regression tests passed.
• The evaluator performs spatial pairing before reading strings, preventing transcription text from driving localization matches.
• The current strict benchmark still requires further adjudication because eight visible-label boxes remain ambiguous.

Results

Selected, bounded observations

57 / 65

confirmed-label localization87.69%; strict current evaluator

89.47%

drawing-wide precisionscope-adjudication based

0

auto-accepted recordsAUTO intentionally disabled

Reported results

• The latest strict confirmed-label evaluation reports 57/65 localization recall (87.69%).
• Drawing-wide precision is 89.47%; the score includes visible valid line labels, duplicates, and scope classification rather than treating every prediction as a register row.
• AUTO acceptance remains disabled with 0 accepted records, even where diagnostic auto-eligibility was observed.

Key takeaways

• Confidence must not substitute for review when OCR alternatives, nearby labels, or broad clusters can change the match.
• A useful engineering-automation MVP can reduce review effort while still making every acceptance decision inspectable.

Limitations

• This is an assisted-review benchmark, not an autonomous line-register system. FROM/TO, connectivity, piping graph inference, production numbering, and final Excel output are outside scope.
• The current strict benchmark requires further adjudication of eight ambiguous visible-label occurrences before a high-stakes performance claim would be appropriate.

What I learned

• Lock predictions before opening truth so evaluation remains auditable.
• Keep geometric localization and text scoring separate, then preserve raw evidence for a reviewer.

Evidence trail