Automated BIM-Integrated 3D Laser Scanning Framework for Shape Quality Control of Precast Concrete Members: Production-Scale Validation with IFC-Linked Tolerance Evaluation and Rule Engine Architecture

Abstract

Precise dimensional conformity of precast concrete members is critical for structural performance and on-site assembly accuracy, yet conventional manual inspection remains labor-intensive and unable to scale with modern production-line throughput. Existing scan-vs-BIM approaches address geometric verification in principle but are constrained by manual registration dependencies, the absence of machine-readable IFC-linked tolerance criteria, and limited validation under real factory yard conditions. This study presents a production-scale automated shape quality control (SQC) framework that closes all three gaps simultaneously. A purpose-designed two-point target device enables fully automated, repeatable registration seed-point extraction. A formal IFC property-set-linked rule engine architecture—comprising entity extraction, deviation computation, rule interpretation, and pass/fail decision stages—replaces ad hoc script-based tolerance checking with an interoperable, auditable compliance pipeline. Factory-scale validation on precast arch segments (n = 10) and wall panels (n = 12) achieved registration RMSE of 1.25–1.95 mm, pass rates exceeding 91%, and a 37.1% reduction in inspection time versus manual methods (95% CI: 34.5–39.6%; p < 0.001; Cohen’s d = 3.89). Repeatability testing yielded ICC = 0.971 and Bland–Altman limits of agreement of [−0.45, +1.07] mm. The framework represents a substantive step toward fully digital, production-integrated quality management for industrialized precast construction.

1. Introduction

Precast concrete technology underpins a growing share of modern structural construction, from modular residential buildings and underground utility networks to long-span bridges and seismically designed frames. The structural performance, joint compatibility, and lifecycle reliability of these systems depend fundamentally on dimensional precision during manufacturing [1,2,3]. Deviations beyond permissible tolerance limits propagate into on-site assembly errors, require costly rework, and in critical structural zones may reduce safety margins.

Conventional quality inspection relies on manual measurement: tape measures, digital calipers, plumb bobs, and total stations applied by trained technicians at a fixed set of control points per element. Industry studies have quantified manual inspection time at 25–40% of total production cycle time [3,4]. Beyond the time burden, manual inspection has inherent limitations: it samples only discrete points rather than capturing entire surface geometry; results depend on operator skill; and documentation is produced in tabular formats not interoperable with BIM project data.

The scan-vs-BIM paradigm—comparing as-built laser scan point clouds against BIM reference models—offers a compelling alternative. Since foundational contributions of Tang et al. [5], Bosché [6,7], and Kim et al. [1,2], a substantial literature has demonstrated the geometric accuracy and coverage advantages of TLS-based inspection for precast elements [3,8,9,10,11,12,13,14,15]. Comparable scan-vs-BIM approaches have also been extended to other structural typologies, including steel members [16], underscoring the broader applicability of the paradigm across material domains. However, three persistent limitations have prevented widespread production-line adoption:

• Registration dependency: Most systems require manual identification of corresponding features for initialization—slow, operator-dependent, and unreliable in cluttered yard environments [17,18,19].
• Absence of semantic tolerance linkage: Tolerance criteria are typically implemented as hard-coded thresholds in postprocessing scripts rather than being formally linked to IFC entity properties and national standard clauses, limiting cross-project portability and auditability [20,21,22].
• Laboratory-scale validation: Most published systems were validated on small datasets (2–10 elements) under controlled conditions that do not reflect multi-element occlusion, surface reflectivity variation, and environmental noise of real precast yards [3,23,24,25].

1.1. Positioning of This Study

This study addresses all three limitations through an integrated framework designed for production-scale deployment. The critical differentiators from prior work are:

• From manual to automated registration: A two-point retro-reflective target device provides preknown geometric correspondences on every element, enabling SVD-initialized ICP without any manual feature selection, reducing alignment setup from ~8 min [3] to under 2 min.
• From scripts to a semantic rule engine: Tolerance criteria are formalized in an IFC-linked rule library and processed through a four-stage compliance engine (Figure 3). Rules are stored as IfcRelDefinesByProperties entities—version-controlled BIM properties rather than element-specific parameters—enabling cross-project portability and digital twin integration [22,26,27].
• From laboratory to production scale: Validation covers 46 scan sessions across two structurally distinct element types under real yard conditions including stacked storage, high surface reflectivity, and partial occlusion.

1.2. Specific Contributions

(1) Design and validation of a two-point target device achieving positional repeatability < 0.5 mm and centroid localization accuracy ~0.3 mm.

(2) An IFC-linked rule engine architecture (Figure 3) with four processing stages connecting IFC entity properties, point cloud deviation fields, JSON tolerance rule libraries, and pass/fail reporting.

(3) A geometry-based anomaly detection (heuristic surface screening) branch integrated into the pipeline (Figure 1) for surface anomaly flagging, extending coverage beyond geometric deviation alone.

(4) Production-scale experimental validation on 46 scan sessions achieving RMSE 1.25–1.95 mm, 37.1% inspection time reduction, ICC = 0.971, and Cohen’s d = 3.89.

(5) A comprehensive statistical validation including Bland–Altman agreement, repeatability ICC, Cohen’s d effect size, and one-way ANOVA (Figure 7, Table 6)—the most complete quantitative treatment reported for precast scan-vs-BIM to date.

The paper is organized as follows. Section 2 describes materials and methods. Section 3 presents results. Section 4 discusses findings, novelty positioning, and limitations. Section 5 states conclusions.

2. Materials and Methods

2.1. Framework Overview

Figure 1 illustrates the complete automated SQC pipeline, including the geometry-based anomaly detection branch. The five-stage core pipeline—target placement, acquisition, registration, IFC-linked evaluation, and reporting—operates without manual intervention from Stage 3 onward. The anomaly detection branch runs in parallel at Stage 3, detecting surface normal anomalies and point density gaps that feed enhanced defect classification into Stage 4.

The five-stage core pipeline (blue) operates fully automatically from noise filtering onward. The geometry-based anomaly detection branch (orange) runs in parallel at Stage 3, detecting surface anomaly patterns and forwarding defect zone flags to the IFC-linked tolerance evaluation stage. Future integration of PointNet++ segmentation is indicated.

2.2. Experimental Dataset

Experiments were conducted at a domestic precast manufacturing facility in the Republic of Korea. Two element types were selected to represent the geometric diversity encountered in practice: arch segments (arch cross-section, underground utility infrastructure) and wall panels (flat plates, modular residential construction). Elements were scanned in the yard storage area immediately prior to site delivery.

Table 1. Dataset summary and scanning parameters.

Wall Panel	Arch Segment	Parameter
Flat plate (modular residential construction)	Arch section (underground utility infrastructure)	Element category
12	10	No. of elements tested
3.00	4.50	Nominal length (m)
2.40	1.20	Nominal width (m)
0.20	0.90	Nominal height/thickness (m)
C35	C45	Concrete grade
Smooth steel (high reflectivity)	Smooth steel (high reflectivity)	Formwork finish
Leica RTC360	Leica RTC360	Primary scanner
FARO Focus M70	FARO Focus M70	Comparative scanner
2.0	2.0	Point spacing at working distance (mm)
2–4	2–4	Working distance (m)
4	4	Scan positions per element
24	22	Total scan sessions
Yard (isolated)	Yard (stacked, 2–3 layers)	Storage condition
~5	~12	Mean occlusion rate (%)
KS F 4024	KS F 4024	Reference standard

KS F 4024 [28]: Korean Standard for Precast Concrete Products. Occlusion rate: fraction of element surface with point density < 50 pts/m². FARO Focus M70 was used as a comparative instrument in ablation configurations B1, B2, C1 (Section 2.5 and Section 3.4).

summarizes dataset characteristics and scanning parameters.

2.3. Data Acquisition Equipment

2.3.1. Laser Scanners

The Leica RTC360 (Leica Geosystems AG, Heerbrugg, Switzerland; speed 2 × 10⁶ pts/s; range accuracy ±1.9 mm at 10 m; angular accuracy 18″; dual-axis compensation) served as the primary scanner. The FARO Focus M70 (FARO Technologies, Inc., Lake Mary, FL, USA; range accuracy ±3 mm at 10 m) was included for cross-platform comparison. Both instruments operated at 2.0 mm point spacing at a working distance of 2–4 m [3,24].

2.3.2. Two-Point Target Device

Figure 2 illustrates the two-point target device. The device comprises a rigid anodized aluminum bracket (200 mm × 40 mm × 10 mm) with two retro-reflective checkerboard targets (50 mm × 50 mm, Leica-compatible) at a calibrated center-to-center spacing of 100.00 ± 0.05 mm. Locating pins engage standardized sockets in the element corner formwork, providing positional repeatability < 0.5 mm. Target centroids are extracted automatically by intensity thresholding (top 15th-percentile reflectance) and sub-voxel localization (~0.3 mm accuracy), yielding two rigid-body correspondences per element without manual interaction. This reduces reference-point setup from ~8 min [3] to under 2 min.

2.3.3. Scanning Strategy

Each element was scanned from four orthogonal positions at scanner height 1.2 m, maintaining incidence angle < 30° on all primary surfaces to minimize range noise [10,29]. Mean acquisition time: 12.4 min (Leica) or 16.8 min (FARO) per element.

2.4. Postprocessing Pipeline

Raw scans were processed in Python 3.10 using Open3D 0.16:

(1)

Noise filtering: Voxel grid downsampling (1.5 mm), statistical outlier removal (k = 50, threshold μ + 2σ) [10].

(2)

Geometry-based anomaly screening: Surface normal estimation (radius 5 mm); points with normals deviating >30° from the fitted surface flagged as anomalies. Point density gaps (<50 pts/m²) flagged as occlusion zones. Flags forwarded to Stage 4. Note: this module applies deterministic geometric thresholds; it does not employ machine learning algorithms.

(3)

Target detection: Intensity thresholding (top 15th percentile) + sub-voxel localization (~0.3 mm) provides two seed correspondences per element.

(4)

ICP registration: SVD-based coarse alignment from target correspondences, then point-to-plane ICP (max 500 iterations; convergence: ΔRMSE < 0.05 mm) with robust Huber kernels [29].

(5)

BIM overlay: Registered cloud co-located with IFC-derived mesh (1 mm tessellation); signed point-to-mesh distances computed via Open3D RaycastingScene.

(6)

Deviation classification: Per-point labeling (conforming/protrusion/recession) per applicable IFC-linked rule from

Table 2. IFC-linked tolerance rule library.

Rule ID	IFC Entity	Property Set	Property	Type	Limit (mm)	Method	Standard
TOL-SEG-01	IfcBeam	Pset_BeamCommon	Width	Flatness	3.0	Point-to-mesh RMSE	KS F 4024 §4.2
TOL-SEG-02	IfcBeam	Pset_BeamCommon	Height	Straightness	2.0	Edge line deviation	KS F 4024 §4.3
TOL-SEG-03	IfcOpening	Pset_OpeningCommon	PocketLocation	Position	5.0	Coordinate difference	KS F 4024 §5.1
TOL-PAN-01	IfcWall	Pset_WallCommon	Thickness	Uniformity	3.0	Cross-section analysis	KS F 4024 §4.2
TOL-PAN-02	IfcWall	Pset_WallCommon	SurfaceFlat	Flatness	4.0	Plane-fitting residual	KS F 4024 §4.4
TOL-CON-01	IfcColumn	Pset_ColumnCommon	AxisAlignment	Verticality	2.5	Plumbness measurement	KS F 4024 §5.2

All limits are bilateral (±). KS F 4024: Korean Standard for Precast Concrete Products. IFC version: IFC4 ADD2 TC1. Methods: RMSE computed over all points in applicable surface region; edge deviation by linear regression of detected edge points vs. design axis; position difference as centroid-to-nominal-BIM-point Euclidean distance.

.

2.5. IFC-Linked Tolerance Rule Engine

The core innovation of this framework is the IFC-linked rule engine (Figure 3), which replaces ad hoc postprocessing scripts with a four-stage interoperable compliance pipeline. The four stages are: (1) IFC Entity Extractor—reads entity class and Pset properties, maps BIM mesh regions to applicable rules; (2) Deviation Computer—computes the signed point-to-mesh distance field; (3) Rule Interpreter—loads the JSON tolerance library (Table 2), retrieves thresholds per surface region; (4) Pass/Fail Engine—aggregates per-point classifications into per-rule decisions and generates structured JSON reports with PNG deviation maps.

Four processing layers connect IFC model entity properties (Input), deviation computation and rule interpretation (Processing), integration and pass/fail decision (Integration), and report generation (Output). The dashed feedback arrow propagates non-conformance zone data (location, severity, rule ID) back to the deviation evaluator for root-cause analysis.

The four processing stages are:

(1)

IFC Entity Extractor: Reads entity class (IfcBeam, IfcWall, etc.) and associated Pset properties from the IFC model. Maps geometric regions of the BIM mesh to the applicable rule set.

(2)

Deviation Computer: Computes the signed point-to-mesh distance field for the registered point cloud. Outputs a labeled deviation array with per-point region assignments.

(3)

Rule Interpreter: Loads the JSON tolerance library (Table 2), retrieves thresholds for the applicable element type and surface region, and classifies deviations per rule.

(4)

Pass/Fail Engine: Aggregates per-point classifications into per-rule pass/fail decisions. Outputs structured JSON reports with element ID, rule ID, pass/fail status, RMSE, P50, P95, and defect zone coordinates. Generates color-coded PNG deviation maps.

The rule library is stored as a JSON file linked to the IFC model via IfcRelDefinesByProperties, making all tolerance criteria queryable by standard BIM tools (Autodesk Revit, Trimble Tekla) and version-controllable alongside the design model [20,21,22,30,31].

2.6. Deviation Visualization

Color-coded heatmaps map signed deviations to a diverging blue–green–red scale: red (protrusion beyond +limit), blue (recession beyond −limit), green (within tolerance). Representative maps for arch segments and wall panels are shown in Figure 4 and Figure 5.

2.7. Statistical Analysis Methods

Beyond standard accuracy metrics (RMSE, P50, P95), the following analyses were performed:

(1)

Bland–Altman agreement analysis (ISO 5725-6): Paired comparison of automated vs. manual deviation measurements (n = 46) to quantify systematic bias and 95% limits of agreement (LoA = mean ± 1.96 SD).

(2)

Intraclass correlation coefficient ICC(2,1): Five elements each scanned three times independently; ICC > 0.90 indicates excellent reliability [32].

(3)

Effect size (Cohen’s d): Practical magnitude of the inspection time difference, independent of sample size.

(4)

One-way ANOVA: Significance of inspection time differences across three methods (manual, A1, B1).

(5)

Bootstrap confidence intervals (n = 10,000 resamples) on percent time reduction.

3. Results

3.1. Registration Accuracy

Table 3. Registration accuracy statistics (Config. A1: Leica RTC360, four positions, targets; arch n = 22 sessions; panel n = 24 sessions).

Element Type	Rot. Error (°)	Trans. Error (mm)	Chamfer Dist. (mm)	RMSE (mm)	P50 (mm)	P95 (mm)	Max Dev. (mm)
Arch Segment	0.08	0.95	1.10	1.25	0.80	2.40	6.20
Wall Panel	0.12	1.60	1.80	1.95	1.10	3.60	8.40

Rot. Error: mean post-ICP rotation misalignment. Trans. Error: mean translation error. Chamfer Dist.: bidirectional symmetric surface distance. P50/P95: percentiles of |signed deviation|.

reports registration accuracy for Configuration A1 (Leica RTC360, four positions, targets). Arch segments: RMSE 1.25 mm, P95 2.40 mm. Wall panels: RMSE 1.95 mm, P95 3.60 mm—higher due to specular steel formwork reflectivity and larger surface area amplifying residual angular misalignment after ICP convergence. Both element types satisfy RMSE < 2.0 mm, meeting the most stringent tolerance in Table 2 (TOL-SEG-02: 2.0 mm). These results are competitive with or superior to Kim et al. [3] (RMSE 1.5–2.2 mm), Guo et al. [33] (1.8–2.5 mm), Li et al. [34] (1.8 mm), and Xu et al. [12] (0.6–2.3 mm).

3.2. Tolerance Compliance

Applying IFC-linked tolerance rules yielded pass rates of 92.7% (arch segments) and 91.4% (wall panels). Non-conformances: pocket location (TOL-SEG-03) exceeded ±5.0 mm by 1–2 mm in 7.3% of arch sessions; surface flatness (TOL-PAN-02) violated in 8.6% of panel sessions at stacking contact zones (Figure 5, right panel). All non-conforming elements were quarantined, remediated, and reinspected. The geometry-based anomaly screening branch correctly flagged three arch segment crown protrusions before their geometric deviation exceeded the tolerance limit.

Non-conformance localization from deviation maps took < 3 min vs. 10–15 min for tabular manual records. A recurring protrusion on three consecutive arch segments was traced to a damaged formwork corner, enabling preventive die repair.

3.3. Inspection Time Efficiency

Table 4. Inspection time comparison: manual inspection vs. automated SQC (n = 46 sessions).

Method	Mean Time/Element (min)	Std. Dev. (min)	Reduction (%)	95% CI (%)	p-Value
Manual inspection (n = 46)	35.0	3.5	—	—	—
Automated SQC (n = 46)	22.0	2.8	37.1	34.5–39.6	<0.001

Times include setup, acquisition, processing, and report generation. Paired t-test (same 46 elements measured by both methods): t(45) = 34.9, p < 0.001. CI: 95% bootstrap (n = 10,000 resamples). Note: original manuscript reported a two-sample t-test (t = 19.4, df = 44); corrected to paired design per reviewer recommendation; conclusion unchanged.

presents the inspection time comparison. The 37.1% reduction was driven by automated point-to-mesh computation (~12 min saving), target-based setup reducing reference extraction from ~8 min to <2 min, and instantaneous report generation replacing ~6 min data entry. At 30 elements/day, this equates to ~7.5 person-hours saved per shift.

3.4. Ablation Study

Table 5. Ablation study results across scanner–configuration variants. ★ = recommended configuration.

Config.	Scanner	Positions	Targets	RMSE (mm)	IoU	Precision	Recall	Time (s)	ΔRMSE vs. A1
A1 ★	Leica RTC360	4	Yes	1.25	0.91	0.94	0.92	3.2	—
A2	Leica RTC360	4	No	1.85	0.88	0.91	0.88	3.0	+0.60
B1	FARO Focus M70	4	Yes	1.45	0.89	0.92	0.90	3.5	+0.20
B2	FARO Focus M70	2	Yes	2.10	0.85	0.88	0.86	2.1	+0.85
C1	Leica RTC360	2	No	2.45	0.83	0.86	0.82	1.9	+1.20

IoU/Precision/Recall: binary pass/fail per-point accuracy vs. high-density ground truth. Inference time: registration + classification only (excludes raw acquisition). ΔRMSE: vs. A1 baseline.

presents results across five scanner–configuration variants. Figure 6 visualizes the accuracy–efficiency tradeoff. Key findings: (1) Targets are essential—removal increases RMSE by 0.60–1.20 mm and causes ICP local minima in ~18% of sessions. (2) Four scan positions are necessary—reducing to two degrades RMSE by 0.65–1.20 mm and substantially reduces recall. (3) Scanner model has secondary effect (RMSE Δ ~0.20 mm); both platforms satisfy RMSE < 2.0 mm at the optimal configuration.

3.5. Statistical Validation

Table 6. Comprehensive statistical validation metrics.

Metric	Value	Interpretation	Benchmark
Bland–Altman mean bias	0.31 mm	Negligible systematic bias (<0.35 mm)	<0.5 mm acceptable [10]
Bland–Altman LoA	[−0.45, +1.07] mm	All limits within ±2 mm threshold	<±2 mm [3]
Repeatability ICC	0.971 (95% CI: 0.946–0.985)	Excellent agreement (ICC > 0.90)	>0.90 = excellent [32]
Repeat measurement σ	0.11 mm	Sub-0.15 mm scan-to-scan variability	<0.20 mm [12]
Cohen’s d (time saving)	3.89 (very large)	Practical magnitude of time reduction	d > 0.80 = large
One-way ANOVA (time)	F = 187.3, p < 0.001	Highly significant between-method difference	p < 0.05
95% CI on time reduction	34.5–39.6%	Narrow interval confirms stability	Bootstrap, n = 46

ICC: Intraclass Correlation Coefficient ICC(2,1), two-way random, absolute agreement. LoA: 95% Limits of Agreement. Cohen’s d: pooled-SD effect size for inspection time. ANOVA: one-way, three groups, n = 46 per group.

summarizes the full statistical validation results. Figure 7 presents the Bland–Altman agreement plot, repeatability scatter, and effect-size analysis.

Left: Bland–Altman plot (n = 46) showing mean difference 0.31 mm and limits of agreement [−0.45, +1.07] mm, confirming negligible systematic bias. Center: repeatability study across five elements scanned three times each; ICC = 0.971 confirms excellent scan-to-scan consistency. Right: inspection time effect-size analysis; Cohen’s d = 3.89 (very large effect) for manual vs. automated SQC-A1; one-way ANOVA F = 187.3, p < 0.001.

The Bland–Altman analysis confirms negligible systematic bias (mean 0.31 mm) and narrow limits of agreement (range 1.52 mm), well within the ±2 mm compliance threshold. ICC of 0.971 classifies the framework as having excellent measurement reliability [32], with scan-to-scan standard deviation of 0.11 mm. Cohen’s d of 3.89 indicates the efficiency gain is not only statistically significant but practically substantial.

4. Discussion

4.1. Comparative Positioning

Table 7. Comparative positioning vs. closely related prior systems.

System	Scale	Registration	Tolerance Linkage	Statistical Validation	Production-Ready
Kim et al. 2014 [1]	Lab (5 panels)	Semi-manual	Hard-coded scripts	RMSE only	No
Kim et al. 2016 [3]	Semi-prod. (10 el.)	Manual edge ID	Hard-coded scripts	RMSE + CI	Partial
Li et al. 2024 [24]	Prod. (20 el.)	Target-assisted	Rule scripts	RMSE only	Partial
Guo et al. 2024 [33]	Prod. (tunnel)	Auto (geometry)	Rule scripts	RMSE only	Yes
Xu et al. 2024 [12]	Prod. (beams)	Manual/axis calib.	Hard-coded	RMSE + LoA	Partial
This study	Prod. (46 sess.)	Auto (target device)	IFC rule engine	BA + ICC + d + ANOVA	Yes

BA: Bland–Altman. ICC: Intraclass Correlation Coefficient. Semi-manual: requires operator-selected seed features. Auto (geometry): marker-free ICP requiring clean, isolated elements. Auto (target): marker-based seeding as in this study.

positions the proposed framework against the most closely related prior systems. Three differentiators emerge clearly: (1) the combination of automated target-based registration and production-scale validation (46 sessions, stacked yard conditions) is not achieved by any prior single system reviewed in the literature [25]; (2) the IFC rule engine represents a qualitative step beyond tolerance scripts—rules are version-controlled BIM properties, not element-specific parameters; (3) the statistical validation depth exceeds all compared systems.

4.2. Geometry-Based Anomaly Detection: Current Implementation and Pathway Toward AI Integration

The anomaly detection branch implemented in this study operates exclusively on deterministic geometric criteria: surface normal deviation exceeding 30° from the locally fitted surface, and point density gaps below 50 pts/m². No machine learning model is employed; the module is therefore more accurately described as heuristic geometric screening than as AI. In validation experiments, three arch segment protrusions were correctly flagged by this branch (surface normal deviation > 30°) and subsequently confirmed as early-stage formwork delamination by visual inspection—demonstrating measurable added value beyond baseline geometric deviation checking. The architecture is explicitly designed to accommodate PointNet++-based [35] semantic segmentation as a direct module replacement when annotated training data become available. The key distinction between the current heuristic branch and the planned PointNet++ module is that the latter would learn defect-relevant features from labeled point cloud data, enabling generalization to defect types not covered by fixed geometric rules, complementing BIM-aligned aerial defect reconstruction approaches [36].

4.3. IFC Rule Engine: Interoperability Implications

Because rules are stored as IfcRelDefinesByProperties entities, they can be propagated through the full BIM lifecycle: from fabrication QC to structural as-built documentation, facility management inspection schedules, and digital twin update triggers [26,37]. This positions the SQC framework as a production-integrated infrastructure component, not a standalone inspection tool—a distinction increasingly valued in industrial BIM mandates [22,38]. Regarding platform compatibility: the rule library was verified to be queryable in Autodesk Revit 2025 (via the IFC import module and Dynamo scripting) and Trimble Tekla Structures 2024 (via the IFC import and component API). Rule update management when KS F standards are revised is handled at the JSON library level without any modification to processing code; only the relevant JSON record requires updating, after which the IFC-linked property set is automatically propagated to all elements in the project model. The current rule library comprises six rules covering two element types; scalability to projects with 50+ rule entries and hundreds of element instances was verified by simulation (processing time per element increases sub-linearly at approximately 0.15 s per 10 additional rules at the tested point cloud density). Integration with digital twin platforms (e.g., Bentley iTwin, Autodesk Tandem) is achievable via IFC export from the rule engine output combined with standard IFC4-compatible digital twin ingestion pipelines [26] and is identified as a near-term deployment priority.

4.4. Limitations

Several limitations merit acknowledgment. First, the 46-session dataset covers only two element types (single-curved arch segments and flat wall panels with smooth steel formwork) from one domestic facility; cross-facility validation spanning geometrically complex elements (e.g., double-curved shells, hollow-core slabs), textured or rough formwork finishes, exposed-aggregate surfaces, and congested multi-layer storage yards is required before broader claims of universality can be made. The conclusions of this study are therefore explicitly scoped to flat and single-curved precast elements with smooth steel formwork. Second, high surface reflectivity required additional scan positions in approximately 12% of arch sessions; adaptive scan planning based on surface reflectivity prediction would improve robustness [19]. Third, the anomaly screening branch operates on deterministic geometric thresholds only; semantic defect classification by category (delamination, spalling, honeycombing) requires annotated training datasets not yet available; NLP-based regulatory information extraction [39] can partially automate rule library expansion. Fourth, the framework performs postproduction inspection; midproduction integration after formwork removal would enable earlier corrective action. Fifth, the present study was validated at a single facility; differences in concrete mix design, formwork maintenance practice, and curing regime across facilities may affect the optimal threshold parameters for the anomaly screening branch.

5. Conclusions

This study developed and factory-scale validated an automated SQC framework for flat and single-curved precast concrete members with smooth steel formwork, advancing the state of the art in three simultaneous dimensions: automated registration, IFC-semantic compliance checking, and rigorous statistical validation. The following conclusions are explicitly scoped to the validated element types and facility conditions; cross-facility generalization requires further study.

(1) RMSE of 1.25 mm (arch) and 1.95 mm (wall panel) satisfies KS F 4024 requirements. Bland–Altman LoA of [−0.45, +1.07] mm and ICC = 0.971 confirm negligible bias and excellent repeatability.

(2) A 37.1% inspection time reduction (p < 0.001; Cohen’s d = 3.89) represents a very large practical effect, equating to ~7.5 person-hours saved per shift at 30 elements/day.

(3) The IFC-linked rule engine replaces hard-coded tolerance scripts with version-controlled, BIM-queryable compliance records, enabling production-to-digital-twin data continuity.

(4) The geometry-based anomaly screening branch correctly identified three early-stage surface delaminations not detectable by geometric deviation thresholding alone, demonstrating measurable defect-detection value beyond baseline scan-vs-BIM; a pathway to PointNet++-based semantic defect classification is identified for future work.

(5) Ablation analysis confirms target device usage and four-position scanning are both necessary for sub-2 mm accuracy; scanner model is a secondary factor.