AURA: Additive Manufacturing Utility for Real-time Anomaly Detection

Framework for LPBF Thin-Wall Geometric Monitoring

---

📌 Project Overview

AURA (Additive Manufacturing Utility for Real-time Anomaly Detection, evolved from its initial purpose as an AES Utility for Real-time Anomaly Detection) is a lightweight, modular framework for detecting geometric anomalies in Laser Powder Bed Fusion (LPBF) thin-wall structures through multi-modal optical and temporal analysis. The framework monitors 316L stainless steel thin walls (0.2–1.0 mm) during manufacturing using layer-wise optical imaging for in-situ wall thickness, surface roughness, and center drift quantification [1].

A key innovation is repurposing "timing" models originally designed for AES-128 side-channel analysis to analyze temporal dynamics (layer-to-layer evolution), enabling robust detection of manufacturing instabilities. The framework supports:

• Spatial detection — Wall thickness deviation, surface roughness, and center drift from individual layer images
• Temporal detection — Layer-to-layer cumulative changes indicating process instability
• Hybrid detection — Fusion of spatial + temporal features for improved robustness

The framework implements three primary detection strategies:

• Threshold-based detection — Statistical anomaly identification using adaptive thresholds on geometric features
• ML-based detection — Random Forest classifiers trained on engineered geometric features
• Hybrid detection — Fusion of both spatial and temporal detection approaches

---

⚙️ Features

• 🔍 LPBF In-Situ Monitoring — Layer-by-layer geometric tracking during thin-wall fabrication based on reference [1]; adapted from original AES-128 side-channel monitoring
• 📊 Multi-Modal Feature Extraction (per layer image):
  • Spatial features: Wall thickness (px), center X position, left/right edge detection, surface roughness (Laplacian), texture variance, contour irregularity
  • Temporal features: Layer-to-layer thickness deviation (%), center drift (px), rolling statistics, accumulation metrics, thickness velocity (evolved from AES execution timing patterns)
  • Hybrid vectors: Combined spatial + temporal for fused detection
• 👁️ Computer Vision Pipeline:
  • Gaussian blur + horizontal gradient-based edge detection (per reference [1])
  • Intensity profile averaging for robust wall boundary identification
  • Multi-scale roughness analysis via Laplacian operator
• 🎯 Three Detection Modes (originally developed for AES timing analysis, adapted for LPBF):
  • Spatial Detection — Threshold-based anomaly detection on wall geometry
  • Temporal Detection — Threshold-based anomaly detection on layer-to-layer changes
  • Hybrid Detection — Fusion of spatial + temporal criteria for improved recall
• 🤖 ML-Based Variants — Random Forest classifiers for all three modes + adaptive thresholding (maintains original AES model architecture)
• 📷 Batch Image Processing — Efficient frame-by-frame feature extraction from layer image directories
• 📈 Automated Reporting — Excel (.xlsx) export with detailed geometric metrics, anomaly flags, and statistics
• 🏃 Lightweight & Fast — ~100ms per layer inference, suitable for real-time manufacturing feedback

---

📂 Repository Structure

AURA/
├── README.md                    # Project documentation
├── Requirements.txt             # Python dependencies
├── ML_Detect.py                 # [Legacy] AES-based ML detector (kept for compatibility)
├── Threshold_Detect.py          # [Legacy] AES-based threshold detector (kept for compatibility)
├── thermal_adapter.py           # [Legacy] Signal adapter (kept for compatibility)
│
├── data/                        # LPBF layer image datasets
│   ├── first layer/             # Optical images from Layer 1 (42 frames)
│   ├── second + third/          # Optical images from Layers 2-3
│   └── layer1.png - layer7.png  # Individual layer snapshots
│
└── Testing/                     # LPBF geometric monitoring test suite
    ├── _utils.py                # Shared utilities: image loading, geometric feature extraction per [1]
    ├── run_all.py               # Orchestrator for running all 6 test variants
    │
    ├── spatial_threshold.py      # [=optical_threshold.py] Spatial geometric anomaly detection
    ├── spatial_ml.py            # [=optical_ml.py] Spatial ML anomaly detection
    │
    ├── temporal_threshold.py     # [=timing_threshold.py] Temporal (layer-to-layer) anomaly detection
    ├── temporal_ml.py           # [=timing_ml.py] Temporal ML anomaly detection
    │
    ├── hybrid_threshold.py       # [=timing_optical_threshold.py] Spatial+temporal threshold fusion
    └── hybrid_ml.py             # [=timing_optical_ml.py] Spatial+temporal ML fusion

Note: File naming convention (optical_* / timing_*) is retained for backward compatibility with prior ablation study naming. Semantically:

• optical_* files = Spatial detection (wall geometry from images)
• timing_* files = Temporal detection (layer-to-layer dynamics)
• timing_optical_* files = Hybrid detection (spatial + temporal fusion)

Core Framework Files (Root Directory)

The three primary files maintain backward compatibility with AES framework:

File	Status	Purpose
ML_Detect.py	Legacy	AES-based ML anomaly classifier (replaced by Testing/spatial_ml.py for LPBF)
Threshold_Detect.py	Legacy	AES-based threshold detector (replaced by Testing/spatial_threshold.py for LPBF)
thermal_adapter.py	Legacy	Signal adapter (still usable with Testing models)

For LPBF thin-wall monitoring, use the Testing suite directly.

Testing Suite (Testing/ Directory)

Six LPBF geometric monitoring test scripts organized by detection strategy:

Detection Strategy	Threshold-Based	ML-Based
Spatial Only (from layer images)	optical_threshold.py	optical_ml.py
Temporal Only (layer-to-layer metrics)	timing_threshold.py	timing_ml.py
Hybrid (spatial + temporal fusion)	timing_optical_threshold.py	timing_optical_ml.py

---

🚀 Setup Instructions

Prerequisites

• Python 3.7+
• Virtual environment (recommended: venv or conda)
• Dependencies: pandas, numpy, scikit-learn, opencv-python, openpyxl, psutil
  • Note: pycryptodome is optional (legacy AES framework only)

Installation

1. Clone the Repository:

   git clone https://github.com/nishant640/AURA.git
   cd AURA

2. Create and Activate Virtual Environment:

   python3 -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install Dependencies:

   pip install -r Requirements.txt

---

🧪 Usage

LPBF Geometric Monitoring (Current Framework)

Running All Tests with Orchestrator

Automatically run all 6 test variants on LPBF layer images and generate consolidated metrics:

python3 Testing/run_all.py <image_directory> [thickness_threshold] [run_label]

Example:

python3 Testing/run_all.py "./data/first layer" 5 lpbf_first_layer

Arguments:

• image_directory: Directory containing layer image files (.png, .jpg, .bmp, etc.)
• thickness_threshold: Maximum acceptable wall thickness variance (%, default: 5)
  • Threshold defines anomaly as: thickness deviation > ±thickness_threshold%
  • Lower values = stricter anomaly detection
  • Default 5% suitable for 316L stainless steel thin walls
• run_label: Experiment identifier (default: auto-generated timestamp)

Output:

• reports/<run_label>/experiment_summary.txt — Consolidated comparison table (all 6 models)
• reports/<run_label>/metrics.json — Machine-readable metrics
• reports/<run_label>/*.log — Individual test logs with detailed feature extraction

Sample Output Metrics (experiment_summary.txt):

AURA LPBF Geometric Monitoring Experiment Report
================================================================================
Run label: lpbf_first_layer
Image directory: /Users/mcrye/Downloads/AURA/data/first layer
Thickness threshold: ±5%
Started: 2026-04-17 19:20:00 UTC


Running: optical_threshold.py
--------------------------------------------------------------------------------
  Exit code: 0
  Log file: reports/lpbf_first_layer/optical_threshold.log

  Metrics:
    - memory_mb: 274.38
    - detected_anomalies: 39
    - tp: 39
    - fp: 0
    - fn: 0
    - accuracy: 100.00
    - precision: 100.00
    - recall: 100.00


Running: optical_ml.py
--------------------------------------------------------------------------------
  Exit code: 0
  Log file: reports/lpbf_first_layer/optical_ml.log

  Metrics:
    - memory_mb: 140.13
    - actual_anomalies: 8
    - detected_anomalies: 9
    - tp: 8
    - fp: 1
    - fn: 0
    - accuracy: 97.62
    - precision: 88.89
    - recall: 100.00
    - f1: 94.12


Running: timing_threshold.py
--------------------------------------------------------------------------------
  Exit code: 0
  Log file: reports/lpbf_first_layer/timing_threshold.log

  Metrics:
    - memory_mb: 228.50
    - detected_anomalies: 9
    - tp: 9
    - fp: 0
    - fn: 0
    - accuracy: 100.00
    - precision: 100.00
    - recall: 100.00


Running: timing_ml.py
--------------------------------------------------------------------------------
  Exit code: 0
  Log file: reports/lpbf_first_layer/timing_ml.log

  Metrics:
    - memory_mb: 45.04
    - actual_anomalies: 8
    - detected_anomalies: 9
    - tp: 8
    - fp: 1
    - fn: 0
    - accuracy: 97.62
    - precision: 88.89
    - recall: 100.00
    - f1: 94.12


Running: timing_optical_threshold.py
--------------------------------------------------------------------------------
  Exit code: 0
  Log file: reports/lpbf_first_layer/timing_optical_threshold.log

  Metrics:
    - memory_mb: 113.28
    - detected_anomalies: 9
    - tp: 9
    - fp: 0
    - fn: 0
    - accuracy: 100.00
    - precision: 100.00
    - recall: 100.00


Running: timing_optical_ml.py
--------------------------------------------------------------------------------
  Exit code: 0
  Log file: reports/lpbf_first_layer/timing_optical_ml.log

  Metrics:
    - memory_mb: 82.46
    - actual_anomalies: 8
    - detected_anomalies: 9
    - tp: 8
    - fp: 1
    - fn: 0
    - accuracy: 97.62
    - precision: 88.89
    - recall: 100.00
    - f1: 94.12



================================================================================
CONSOLIDATED METRICS COMPARISON
================================================================================

Detection Performance:
  Model                             Accuracy     Precision    Recall      
  -----------------------------------------------------------------------
  optical_ml                               97.62%      88.89%     100.00%
  optical_threshold                       100.00%     100.00%     100.00%
  timing_ml                                97.62%      88.89%     100.00%
  timing_optical_ml                        97.62%      88.89%     100.00%
  timing_optical_threshold                100.00%     100.00%     100.00%
  timing_threshold                        100.00%     100.00%     100.00%


Anomalies Detected:
  Model                             TP       FP       FN       Total   
  ---------------------------------------------------------------
  optical_ml                          8        1        0        9       
  optical_threshold                   39       0        0        39      
  timing_ml                           8        1        0        9       
  timing_optical_ml                   8        1        0        9       
  timing_optical_threshold            9        0        0        9       
  timing_threshold                    9        0        0        9       


Performance Metrics:
  Model                             Memory (MB)    
  --------------------------------------------------
  optical_ml                                 140.13
  optical_threshold                          274.38
  timing_ml                                   45.04
  timing_optical_ml                           82.46
  timing_optical_threshold                   113.28
  timing_threshold                           228.50


Finished: 2026-04-17 19:28:41 UTC

Running Individual Tests

Spatial (Optical) Anomaly Detection:

python3 Testing/optical_threshold.py ./data/first layer 5
python3 Testing/optical_ml.py ./data/first layer 5

Temporal (Layer-to-Layer) Anomaly Detection:

python3 Testing/timing_threshold.py ./data/first layer 5
python3 Testing/timing_ml.py ./data/first layer 5

Hybrid (Spatial + Temporal) Anomaly Detection:

python3 Testing/timing_optical_threshold.py ./data/first layer 5
python3 Testing/timing_optical_ml.py ./data/first layer 5

Arguments (all scripts):

• Arg 1: Directory containing layer images
• Arg 2: Thickness threshold (±%, e.g., 5 = ±5%)

Outputs (individual script):

• Excel report (e.g., optical_threshold_report.xlsx) with:
  • Per-layer: index, filename, thickness_px, thickness_deviation_%, center_x, center_drift_px, roughness, texture_variance, anomaly_detected (True/False)
  • Summary statistics: mean thickness, std dev, max drift, anomaly count, detection metrics (accuracy, precision, recall)

Understanding the Reports

Geometric Features Explained:

• thickness_px — Wall thickness in pixels (convert to mm by dividing by pixels_per_mm calibration)
• thickness_deviation_% — Deviation from baseline thickness as percentage
• center_drift_px — Displacement of wall center compared to previous layer (pixel offset)
• roughness — Laplacian magnitude (higher = rougher surface)
• texture_variance — Pixel intensity variance (higher = more texture/noise)
• anomaly_detected — Boolean flag indicating threshold-based anomaly or ML prediction

Interpreting Detection Results:

• Spatial Anomaly — Indicates wall thickness, drift, or surface roughness outliers in current layer
• Temporal Anomaly — Indicates abrupt layer-to-layer change (instability onset)
• Hybrid Anomaly — Combined spatial OR temporal criteria triggered (highest recall, lowest false negatives)

Typical Anomaly Signatures (per reference [1]):

• Rapid thickness deviation (>±3%) often precedes failure
• Center drift accumulation >0.1mm indicates geometry degradation
• Roughness spikes correlate with surface oxidation or powder adhesion issues
• Temporal models catch early instability; spatial models catch acute defects

---

Legacy Framework: Direct AES Timing Analysis (Not Updated)

The core framework files (ML_Detect.py, Threshold_Detect.py, thermal_adapter.py) remain available for AES side-channel analysis but are not actively maintained for LPBF use cases. Refer to git history or prior documentation for AES-specific usage.

---

🔧 LPBF Detection Models Breakdown

Spatial (Optical-Only) Models

Purpose: Detect geometric anomalies from individual layer wall images using the methodology from [1]

Features Extracted (per-layer):

• Wall thickness (px), center X position, left/right edge positions

• Surface roughness (Laplacian magnitude), texture variance, contour irregularity

• Thickness deviation (%), center drift (px) vs. rolling baseline

• optical_threshold.py — Adaptive statistical thresholding on wall geometry

  • Detection: thickness_deviation > ±threshold_% OR roughness > baseline + 2σ
  • Threshold: Auto-computed from data (mean + 2-3σ on each feature)
  • Output: Per-layer anomaly flag + detailed geometric metrics
  • Best for: Quick spatial anomaly identification, deployment with minimal training data

• optical_ml.py — Random Forest classifier on engineered geometric features

  • Model: RF (100 trees, max_depth=8, random_state=42)
  • Features: thickness_deviation, center_drift, roughness, texture_variance, contour_irregularity
  • Scaling: StandardScaler on feature magnitudes
  • Output: Feature importance scores + anomaly predictions
  • Best for: Learning spatial defect patterns (e.g., powder adhesion zones)

Temporal (Layer-to-Layer Dynamics) Models

Purpose: Detect manufacturing instability from cumulative layer-by-layer evolution

Features Engineered (temporal):

• Frame-to-frame thickness changes (%), center drift accumulation (px)

• Rolling statistics: thickness_rolling_std (window=3), center_drift_accumulation

• Velocity metrics: thickness % change per layer, drift acceleration

• timing_threshold.py — Detects abrupt layer-to-layer deviations

  • Detection: delta_thickness > threshold OR delta_drift_px > baseline + 2σ OR rolling_std > baseline
  • Rationale: "Timing" reinterpreted as temporal evolution; detects instability onset
  • Output: Per-layer anomaly flag + change magnitude metrics
  • Best for: Early warning of process drift (before acute spatial defects appear)

• timing_ml.py — Random Forest on temporal dynamics features

  • Features: thickness_deviation, center_drift, thickness_rolling_std, thickness_pct_change, drift_accumulation
  • Model: RF (100 trees, max_depth=8)
  • Output: Temporal instability predictions + feature importance
  • Best for: Learning temporal signatures of near-failure conditions

Hybrid (Spatial + Temporal Fusion) Models

Purpose: Combine geometric and temporal signals for robust dual-mode anomaly detection

Detection Logic: Trigger anomaly if EITHER spatial OR temporal criterion met

• timing_optical_threshold.py — Independent thresholds + logical OR fusion

  • Spatial criterion: roughness > threshold OR thickness variance > ±threshold%
  • Temporal criterion: thickness_deviation OR drift_change OR rolling_std
  • Fusion: spatial_anomaly OR temporal_anomaly → alert
  • Reports: Both component anomalies + combined flag
  • Best for: Reducing false negatives (high recall); either sensor type can catch defects

• timing_optical_ml.py — Unified Random Forest on combined 8-feature vector

  • Features: thickness_pct_change, thickness_deviation, center_drift, drift_accumulation, rolling_std, roughness, texture_variance, contour_irregularity
  • Model: RF (100 trees, max_depth=8) trained on hybrid feature space
  • Advantage: ML learns cross-feature interactions (spatial + temporal correlations)
  • Output: Single probabilities with spatial + temporal importance
  • Best for: Highest accuracy; learns subtle multi-modal patterns

Model Selection Guide

Scenario	Recommended Model	Reason
Quick deployment, minimal training	optical_threshold.py	No ML training needed; statistical thresholds from data
Learn spatial defect patterns	optical_ml.py	Identifies texture-based anomalies, powder adhesion zones
Early warning of instability	timing_threshold.py	Detects drift/thickness change before acute defects appear
Detect temporal dynamics	timing_ml.py	Classifies instability signatures (acceleration, jitter)
High recall (catch all defects)	timing_optical_threshold.py	Logical OR: any spatial OR temporal anomaly → alert
Best overall accuracy	timing_optical_ml.py	Learns joint spatial-temporal patterns
Production comparison	run_all.py (all 6 models)	Generates consolidated metrics across all detection modes

---

🧠 Methodology Reference

This framework implements the core in-situ geometric monitoring approach presented in reference [1]:

Core Technique (from [1]):

1. Layer-wise optical imaging (20MP monochrome, ~35mm lens, USB 3.0 at 5 fps)
2. Gaussian blur + horizontal intensity profile gradient analysis
3. Wall boundary detection via positive/negative gradient peaks
4. Per-layer: thickness (px), center position, surface roughness (Laplacian)
5. Temporal: frame-to-frame deltas, rolling statistics, cumulative evolution
6. Anomaly detection: multi-criteria thresholding OR ML classification

Key Findings (316L stainless steel, per [1]):

• Thickness-dependent instability onset: ~0.6–1.0mm wall thickness
• Critical drift signature: >0.1mm cumulative center drift indicates geometry degradation
• Roughness spikes correlate with surface oxidation and powder adhesion
• Temporal prediction: layer-to-layer changes precede acute spatial defects by 2–4 frames

See the References section below for the full paper citation.

---

🛠️ Future Work

• Thicker wall ranges — Extend beyond 316L thin walls to medium-thickness (1–5mm) LPBF parts
• Material expansion — Adapt to other powders (Inconel, Ti6Al4V, aluminum) with re-calibration
• Multi-layer correlation — Develop cross-layer feature dependencies (e.g., Layer N+1 predicted from Layer N trends)
• Real-time API — Streaming detection pipeline for closed-loop process control
• Deep temporal models — LSTM/GRU for long-range instability prediction (5+ frame lookahead)
• Hardware acceleration — GPU-accelerated feature extraction for >10fps processing
• Thermal integration — Fuse optical + thermal (IR camera) for subsurface defect detection
• Porosity correlation — Post-build CT scan validation against detection predictions
• Web dashboard — Real-time monitoring and layer-by-layer anomaly visualization
• Multi-printer validation — Test portability across different LPBF systems (Concept Laser, EOS, SLM Solutions)

---

📖 References

[1] Killedar, D., Adhami, M., Sun, C., Garcia-Sandoval, M. E., Downey, A. R. J., Fu, Y., Yuan, L., and Parikh, Y., "In-Situ Layer-Wise Geometry Extraction of Thin-Wall 316L Structures Fabricated by Laser Powder Bed Fusion," Proceedings of the ASME 2025 Annual Review of Progress in Quantitative Nondestructive Evaluation QNDE2026, paper QNDE2026-191357, September 13–16, 2026, Albuquerque, NM.

@inproceedings{Killedar2026,
  author={Killedar, Digambar and Adhami, Mumin and Sun, Can and Garcia-Sandoval, Mateo E. and Downey, Austin R. J. and Fu, Yanzhou and Yuan, Lang and Parikh, Yash},
  title={In-Situ Layer-Wise Geometry Extraction of Thin-Wall 316L Structures Fabricated by Laser Powder Bed Fusion},
  booktitle={Proceedings of the ASME 2025 Annual Review of Progress in Quantitative Nondestructive Evaluation QNDE2026},
  pages={QNDE2026-191357},
  address={Albuquerque, NM},
  month={September 13--16},
  year={2026}
}

---

🤝 Acknowledgments

This work was supported under the McNair Junior Fellowship and Magellan Scholar Program at the University of South Carolina.

Special thanks to Rye Stahle-Smith for hardware testing and experimental support on LPBF systems.

Laser Powder Bed Fusion experiments performed on [LPBF System Model] at the USC Advanced Manufacturing Lab.

---

🎓 Credits

Framework Adaptation for LPBF: Nishant Chinnasami (2026)
Original AES Framework: Nishant Chinnasami
Advisor: Dr. Rasha Karakchi
Institution: University of South Carolina
Lab: Advanced Manufacturing & Materials Characterization Lab