Measurement of construction projects AI

PROJECT SPECIFICATION

1) Purpose and scope of the project

Purpose: To build an AI-based web application that automatically performs measurements from project documentation (architecture, construction, installations/industries, pzt), combines drawings into a whole, identifies elements (e.g., walls, windows, doors, plasters, structural and installation elements), and then presents a complete measurement in tables with the possibility of visual verification (2D/3D) and pricing.

To start:

• Support for vector files: *.dwg, *.pdf (vector)
• Scope of industries: architecture, construction, installations (HVAC, plumbing, electrical), pzt
• Functions: automatic combining of documentation – for application needs, recognition of elements and measurements, visualization, manual correction option, adding prices and creating preliminary estimates

2) Main product assumptions

1. Uploading: the user uploads a package of DWG/PDF files – vector files (plans, sections, elevations, descriptions, legends; possible backgrounds and XREFs).
2. Automatic processing: the system parses files, normalizes layers, units, and scales, combines and fuses drawings into a coherent reference model.
3. AI recognizes elements: recognizes all elements in the projects
4. Measurement: calculates quantities according to measurement rules (lengths, areas, volumes, pieces); takes into account openings, thicknesses, layers – measurements according to global rules
5. Verification: tabular result + 2D/3D preview with highlighting of counted elements; "compare/explain" mode for ambiguities (conflicts between sheets, missing dimensions, inconsistent scales).
6. Manual correction: editable classifications and geometry (drawing/editing polylines/openings), acceptance/rejection of elements.
7. Prices and estimates: attaching price items and rates (unit price, markups, VAT), export CSV/XLSX/BC3/PDF.
8. Audit and versioning: change history, comparison of measurements, differences between versions of documentation

3) User flow

1. New project → name, investor, location, default units and currencies
2. File upload → drag&drop, consistency check
3. Processing → parsing, normalization, fusion, AI classification, measurement.
4. Review → tables + 2D/3D view; ambiguity panel; filters (industry, floor, layer, type).
5. Correction → geometry editor, change class, combine/split elements, acceptance.
6. Prices → assignment of items, rates, markups.
7. Export/report → XLSX/CSV/BC3, PDF summaries, ambiguity report.

5) Functional requirements (detailed)

5.1 Import and validation of files

• Support: vector files: DWG, DXF (optional), vector PDF
• Automatic detection of scale: from viewports (DWG), from title block (OCR), from markings (e.g., 1:100), from dimensions (DIMSTYLE / recognition of dimension text).
• Normalization of units (mm/m) and reference system (translations, rotations, local systems, UCS).
• Detection of deficiencies and errors

5.2 Fusion of documentation (combining into a whole)

• Grouping of sheets: floors, sections, elevations, industries.
• Registration and alignment of plans/sections (matching axis markers, grids, levels; total 2D/3D transformation).
• Combining architecture with construction

5.3 Recognition of elements (AI + rules)

Requirement: Recognition of elements cannot rely solely on layer names. Layer names are treated only as low-weight auxiliary signals.

Implementation rules:

• Classification decisions are primarily based on geometry and topology – based on machine learning
• In case of conflicting information: priority: DIM/dimensions > unambiguous textual description > legend/diagram > geometric features > layer name.
• The system must operate with any/non-standard layer naming (different languages, abbreviations, lack of standards). For testing, complete anonymization of layer names is possible.
• The AI/rules component should allow disabling the influence of layers (feature flag) – the result must still be correct.

5.4 Measurement - measurement based on accepted rules and recognition of elements

5.5 Verification and editing

• View 2D (plan, section) and 3D (simple extruded model) with highlighting of counted elements.
• Check mode: click on an element → measurement details + sources (sheets, descriptions), tolerance/uncertainty.
• Editing tools: drawing/editing polylines, moving/scaling, changing class, combining/splitting.
• List of ambiguities: questions to the user, proposed solutions, acceptance/rejection.

5.6 Prices and estimates

• Price dictionaries/items: global item database (for the system) and local for the user – the user can change the local price)
• Mapping: element type → formula → price item.
• Formulas: price = quantity × rate ± markups (labor, materials, equipment, profit, overhead, VAT).
• Versioning of price lists, currencies, price indexing (inflation).
• Export: XLSX/ (Estimate), PDF summaries.

5.7 Exports and integrations

• XLSX/CSV/BC3, PDF reports (screenshots of visualizations + tables).

6) Non-functional requirements

• Measurement accuracy: geometric error ≤ ±2 mm at correct scales;
• Performance: project up to 200 sheets/2 GB – full processing ≤ 20 min; plan for preview ≤ 3 s.
• Scalability: horizontal (task queues), storage for files up to 10 GB/project.
• Security: GDPR, encryption at-rest and in-transit, IAM, audit log, WAF, backups.
• Reliability: processing SLO 99.5%/month; retry/compensations, versioning of models.

8) Processing pipeline (technical steps)

1. Upload & validation → AV scan, metadata extraction, XREF detection.
2. Parsing:
   • DWG/DXF: reading layers, block references, geometry, DIMSTYLE, viewport scale, Model/Paper Space.
   • PDF (vector): paths, polylines, texts (TrueType), layers; OCR only for vector texts as a fallback.
3. Normalization of scale and units: from viewports and titles; automatic matching; flags in case of conflicts.
4. Registration of sheets: matching axes/grids/markers (ICP/feature matching on vectors), 2D/3D transformations.
5. Extraction of semantic objects:
   • Layer rules (e.g., A-WALL, K-…); name templates; dictionary classification.
   • AI on vector graph: polylines → walls (parallel lines), hatch → fills/floors, blocks → equipment.
   • Texts and legends → parsing parameters (thicknesses, materials, types, room numbers).
6. Fusion: conflict resolution (arch vs const), hierarchy of reliability (DIM > description > legend > layer), tracking data provenance.
7. Measurement: rules, units, deductions for openings, aggregations by floors/rooms.
8. Detection of ambiguities: missing parameters, inconsistent scales, conflicting thicknesses – generating questions.
9. Visualization: generation of backgrounds, overlay of objects, simplified 3D (extrusion according to floor height).
10. Estimates: mapping element→item, formulas, markups, export.

9) AI model and learning

• Representation: vector graph (nodes: geometry vertices and blocks; edges: segments/polylines; attributes: layer, type, line style, nearby texts), additionally rasterization for contextual features.
• Architecture: Graph Neural Network (GAT/GraphSAGE) + Transformer on sequences of vectors; late fusion with CAD rules and heuristics.
• Training data: anchored in real DWG (semi-automatic annotation) + synthetic drawing generators; active learning from user interaction (feedback in UI).
• OCR/NLP: recognition of descriptions, tables, legends; normalization of units (mm/m), parsing expressions (e.g., “wall 24 cm, gypsum plaster 1.5 cm”).
• Metrics: mAP/F1 for classes, MAE/SMAPE for quantities, processing time.
• MLOps: model registry, dataset versioning, validations, A/B in the background per project.

10) Measurement rules – details (configurable examples)

• Walls:
  • Length from the wall axis; Height: from sections / default from the floor; Thickness: from description/layer.
  • Plaster area = 2 × (length × height) – deductions for openings > t_deduction; threshold rules (e.g., do not deduct < 0.5 m²).
• Openings: Width, height from dimensions/markings; number of pieces; area in light.
• Ceilings/slabs: Area from the contour; volume = area × thickness; perimeter (formwork).
• Stairs: Number of runs, area of steps (anti-slip), risers (mb), railings (mb).
• Floors/coverings: Net area after deductions; skirting boards (mb along walls).
• Installations: Lengths of routes along the axis; fittings – detection rules from blocks/symbol dictionary.

Each rule has: quantity definition, data sources, formula, tolerance and priority.

11) UI/UX – key screens

• Project dashboard: processing statuses, latest measurements.
• Documentation review: list of sheets (thumbnails), layers, filters.
• Verification view: 2D/3D canvas with overlay; color legends for classes; element detail panel.
• Measurement table: grouping by industry/floor/room/type; in-line edit; change tracking.
• Ambiguity panel: questions, proposed solutions, acceptance.
• Price lists and mapping: overview of items, formulas, preview of cost impact.
• Exports: range and format selection wizard.