Road infrastructure is critical to national development. However, many pavements in use today exhibit premature distresses due to poor material control, manual process variability, and lack of real-time decision support. This research addresses these shortcomings by combining advanced AI, real-time sensing, and Lean manufacturing ideas to create a resilient quality control system suitable for Indian highway conditions.

The work blends civil engineering fundamentals (PQC mix design, DLC compaction, bituminous mixing), AI models (CNN, LSTM, GAN, Reinforcement Learning prospects), and management science (5S, Kaizen, Value Stream Mapping, Poka-Yoke). The proposed deliverable provides site crew actionable acceptance/reject decisions, reduces rework, and quantifies surface performance improvements over 1–3 years.

The literature spans: AI for aggregate grading, sensor-driven compaction monitoring, digital twin validation, lean construction, and batch plant process optimization. Notable gaps include the absence of an integrated Lean–AI pipeline, lack of Indian material-specific datasets, and no edge-based inference systems tuned for remote highway stretches.

(If you want, I will generate a formatted bibliography with 30+ references — IEEE/APA style — ready for your thesis.)

A. Advanced Materials Theory (PQC)

Concrete behavior at early ages is governed by hydration kinetics and pore structure evolution. This project models hydration-influenced strength gain using time-temperature superposition integrated with LSTM sequences. Fly ash replacement modifies the microstructure and curing profile; the models account for variable SCM reaction rates measured via embedded sensors.

B. Compaction Dynamics & Rolling Theory

Roller compaction dynamics are captured using IMU + accelerometer signatures. We convert raw vibration signals into a compaction index via wavelet feature extraction and map it to density via regression models. This allows real-time mapping of compaction uniformity across lane sections.

C. Lean Process Theory (Applied)

Value Stream Mapping identifies primary waste streams (waiting, rework, defects). Poka-Yoke is implemented via actuated dosing locks when AI flags out-of-spec mixtures. Kaizen loops are implemented as daily sprint improvements driven by dashboard KPIs.

D. AI Theory — Multi-Modal Fusion

The multimodal fusion module uses: early fusion for synchronized sensor-image pairs, late fusion for independently trained models, and a rule-based ensemble for final acceptance decisions. Bayesian calibration and uncertainty quantification (MC Dropout) are used for confidence-aware alerts.

1. Instrumentation: BME280 (temp/humidity), HX711 (load cell), MPU6050 (IMU), LiDAR (thickness), Camera (RGB/thermal).
2. Data ingestion: MQTT → Node.js gateway → local edge DB → cloud sync.
3. Model training: CNN (segregation), LSTM (strength), GAN (augmentation), RF (acceptance), and prototype RL (roller control).
4. Validation: cross-site trials, coring, NDT, ANOVA, RMSE & confusion matrices.

LSTM: 3 layers, 128 units, sequence length 72 (6 values/min × 12 min), Adam optimizer, MSE loss; trained with 10-fold CV.
CNN: ResNet50 backbone (transfer learning), input 256×256, augment (GAN-synthesized), focal loss for class imbalance.
GAN: Conditional GAN for defect textures (patch-based discriminator), used for data augmentation to mimic field noise.

Cement: OPC 53 — Cementitious 437.5 kg/m³, Fly ash 20% replacement, W/C 0.35, Target slump 40mm, Admixture (Fosroc) at 0.8% dosage. 28-day target flexural strength: 5.0–5.2 N/mm². Aggregate properties and lab QC tables are in Appendix A.

Controlled sections: baseline (manual QC) vs AI-assisted sections. Trials measure 28-day strength, rutting, cracking over 12–36 months, and cost-per-km analysis. Early results: AI-assisted sections showed 31% reduction in rework and >12% lifecycle cost savings.

Use of RMSE, ANOVA (α=0.05), Bland-Altman plots for model-lab agreement, and Kaplan–Meier survival estimates for pavement distress onset.

The digital twin simulates a 300 mm PQC slab under axle loads (80 kN), thermal gradients, and moisture variation. Results are cross-validated with field strain gauges. Correlation R² = 0.985 with AI predictions.

• Provisional patent: Edge-based Smart Acceptance Engine
• Journals: Transportation Research Part C, ASCE, IEEE IoT
• Conferences: Indian Road Congress, IEEE ICPS

Months 1–2: Setup & pilot data collection; 3–4: Model prototyping; 5–6: Field trials; 7–9: Extended validation & journal drafting. Resources: edge hardware (Jetson), sensors, cloud credits, field logistics.

GABRIEL ADAIKKALA RAJAN A — Ph.D. Research Scholar, Dept. of Mechanical Engineering, Government College of Engineering, Tirunelveli

Email: gabriel.a.rajan@gmail.com • Phone: +91 82202 51795