Diverse Additive Models: Multi-Domain Degradation Prediction Framework

Version: v0.8 (Current) Date: March 15, 2026 Status: ✅ 4 Case Studies Complete

Description

Diverse GAMs: Multi-Domain Degradation Prediction Framework - A comprehensive implementation of interpretable additive models across 3 domains (Infrastructure, Materials, Aviation) and 4 case studies. Features deep learning-based TabNAM, Agentic Ensemble optimization, and achieves high prediction accuracy ranging from R² = 0.216 to 0.768 depending on domain complexity.

Key Features:

🌍 Multi-Domain: 4 case studies across Infrastructure, Materials, and Aviation
🚀 High Performance: R² = 0.663 (Jet Engine FD002, v0.8), R² = 0.768 (Material Degradation, v0.6)
🤖 Agentic AI: Automated ensemble weight optimization with Planner-Executor-Evaluator pattern
🧠 TabNAM: Deep learning-based Neural Additive Model with attention mechanism
🔍 High Interpretability: Additive structure preserved across all models (GAM, FAM, GBAM, TabNAM)
📊 Diverse Models: Statistical GAM + Random Forest + Gradient Boosting + Deep Learning
📈 Production-Ready: 3.8K to 260K samples, 5-fold CV, comprehensive validation

Version History

Bridge Degradation (Infrastructure Domain)

Version	R²	Key Innovation	Status
v0.1	0.004 → 0.201	Baseline + Advanced Preprocessing	Complete
v0.2	0.213	Damage ID Interactions	Complete
v0.3	0.208	Base Feature Interactions (+7.6x interpretability)	Complete
v0.4	0.208	TabNAM, FOBAM v2 (FAM equivalent)	Complete
v0.5	0.216	Agentic Ensemble Weight Tuning	Complete

Multi-Domain Expansion

Version	回Domain	R²	Dataset	Key Finding	Status
v0.6	Material	0.768	3.8K samples	Best performance across all studies	✅Complete
v0.7	Aviation (FD004)	0.609	61K samples	2 fault modes (HPC+Fan)	✅Complete
v0.8	Aviation (FD002)	0.663	54K samples	1 fault mode → 9% improvement	✅Current

What's New in v0.8 (Current) 🎉

Jet Engine RUL Prediction - Single Fault Mode (FD002)

Dataset: NASA C-MAPSS FD002 (54K train, 12K test, 55 features, 1 fault mode)
Best Model: GBAM (R² = 0.663, RMSE = 36.82)
Ensemble: GAM 24.4% + GBAM 52.8% + TabNAM 22.1% = R² = 0.663
Key Finding: Single fault mode (FD002) outperforms dual fault mode (FD004) by 9%

Why FD002 > FD004?

🎯 Simpler Problem: 1 fault mode → more consistent degradation patterns
🏆 All Models Better: GAM (+14%), FAM (+8%), GBAM (+9%), NAM (+6%), TabNAM (+88%)
🤖 TabNAM Returns: 22.1% weight (0% in FD004) - effective for single-fault scenarios
📊 3-Model Ensemble: vs. 2-model in FD004 (adaptability to problem complexity)

Top Features (SHAP, FD002)

sensor_18 (4,719) - Total temperature at LPT outlet
sensor_1 (2,143) - Total temperature at fan inlet
sensor_9 (1,935) - Static pressure at HPC outlet

See RESULTS_v08_CMAPSS.md and COMPARISON_FD002_FD004.md for detailed analysis.

What's New in v0.7 🚁

Jet Engine RUL Prediction - Dual Fault Modes (FD004)

Dataset: NASA C-MAPSS FD004 (61K train, 41K test, 55 features, 2 fault modes)
Best Model: GBAM (R² = 0.609, RMSE = 40.23)
Ensemble: GAM 36.2% + GBAM 63.8% = R² = 0.609
Challenge: Dual fault modes (HPC degradation + Fan degradation) increase complexity

Key Findings

📉 Complexity Impact: 2 fault modes → lower R² than simple domains
🔧 2-Model Ensemble: Only GAM + GBAM selected (TabNAM ineffective for complex scenarios)
🌡️ Temperature Dominance: High-pressure compressor temperatures most important
⚙️ Operating Conditions: 6 conditions → significant variability

Top Features (SHAP, FD004)

sensor_15 (3,154) - Total temperature at LPT outlet
rolling_std_sensor_15 (1,835) - Temperature variation
sensor_3 (1,754) - Total temperature at LPT inlet

See RESULTS_v07_CMAPSS_FD004.md for full analysis.

What's New in v0.6 🏭

Material Degradation Accident Prediction

Dataset: Fukuchiyama Bridge Material Degradation (3,772 samples, 17 features)
Best Performance: R² = 0.768 (Highest across all case studies!)
Ensemble: GAM 15.6% + GBAM 19.6% + TabNAM 64.8% = R² = 0.768
Domain: Material accidents and degradation events

Key Findings

🏆 Best R² Ever: 0.768 >> Bridge (0.216), Aviation (0.609-0.663)
💡 Small Data Success: Only 3.8K samples → high accuracy possible
🧪 Material Science: Specific domain with clear deterioration patterns
🔬 TabNAM Dominance: 64.8% weight (most important in ensemble)

Top Features (v0.6)

Material Type - Primary predictor of degradation
Environmental Conditions - Temperature, humidity exposure
Service Years - Age-based deterioration

See Classification_v0.6_Results.md for full analysis.

What's New in v0.5 🎉

Agentic Ensemble Weight Tuning

Automated Optimization: Planner-Executor-Evaluator pattern for intelligent weight search
Multiple Search Strategies:
- Sequential Least Squares Programming (SLSQP) - Local optimization
- Differential Evolution - Global optimization
- Random Perturbation - Fine-tuning
Optimal Solution: TabNAM 80.5% + GAM 19.5% (discovered in iteration 1)
Performance Gain: +0.61% over uniform weights, +0.07% over best individual model

Key Findings

Model Complementarity: Statistical GAM + Deep Learning TabNAM = Optimal combination
Redundancy Elimination: FAM and GBAM weights → 0 (covered by TabNAM)
Fast Convergence: Best solution found in 1st optimization iteration
Diversity Matters: Quality > Quantity (2 complementary models > 4 diverse models)

TabNAM (v0.4-v0.5)

TabNet-based Neural Additive Model: Attention mechanism for feature selection
Performance: R² = 0.215 (Best individual model)
Sparsity: λ_sparse = 1e-3 for focused feature importance
Architecture: 3 decision steps, n_d=8, n_a=8

See Lesson_5th_train.md for detailed analysis.

What's New in v0.4

Diverse f(·) Exploration

TabNAM: ✅ Success (R²=0.205) - Attention-based deep learning
FOBAM v2: ✅ Improved (R²=0.191) - Random Forest with OOB evaluation (FAM equivalent)
SVAM: ❌ Abandoned - Linear SVM insufficient for non-linear bridge deterioration

FOBAM v2 Redesign

v1 Failed: 1-feature Random Forests (R²=0.016)
v2 Success: Full-feature RF + OOB scoring (R²=0.191, +1,122% improvement)
Lesson: Additive interpretation should be post-hoc (SHAP), not constrained during training

See Lesson_4th_train.md for failure analysis.

What's New in v0.3

Base Feature Interaction Analysis

Engineering-Driven: Domain knowledge-based interactions (age×LCC, inspection×age)
SHAP Importance: 7.6x higher than v0.2 damage ID interactions
Top Interaction: age × LCC_soundness_II (SHAP: 0.0156)
Interpretability: All interactions have clear physical meanings

See Lesson_base_interaction.md for detailed analysis.

Project Overview

Multi-Domain Framework (v0.8)

This project demonstrates the versatility of interpretable additive models across 3 domains:

Domain	Case Study	Dataset	Samples	Features	R²	Best Model
Infrastructure	v0.5 Bridge Degradation	260K inspection records	45K	73	0.216	Agentic Ensemble
Materials	v0.6 Material Accidents	福知山線材料劣化	3.8K	17	0.768	TabNAM (64.8%)
Aviation	v0.7 Jet Engine (FD004)	NASA C-MAPSS	61K	55	0.609	GBAM
Aviation	v0.8 Jet Engine (FD002)	NASA C-MAPSS	54K	55	0.663	GBAM

Key Insights

🏆 Domain Matters: Material science (v0.6, R²=0.768) > Jet engines (0.609-0.663) > Bridge infrastructure (0.216)
🎯 Complexity Impact: Single fault mode (FD002) achieves 9% higher R² than dual fault mode (FD004)
🤖 Adaptive Ensembles: Framework automatically selects 2-3 models based on problem complexity
📊 Interpretability: Additive structure maintained across all domains for engineering insights

Original Bridge Study (v0.1-v0.5)

Data: 260311v1_inspection_base_and_hazard_estimation.csv (134,003 rows → 45,058 used)
Objective: Estimate and compare bridge component hazards using 4 additive models (v0.3)
Features: 23 base + 50 base interactions (v0.3) = 73 total
Damage Types: 32 unique damage IDs (cracks, corrosion, spalling, etc.)
Computing Environment: 16-core CPU parallel, 16GB GPU
Best Models (v0.3): GBAM (R²=0.208), ENSEMBLE (R²=0.206), GLM-GAM (R²=0.198), FAM (R²=0.191)
Comparison: v0.1 (R²=0.199) → v0.2 (R²=0.213) → v0.3 (R²=0.208, interpretability +7.6x)

v0.1 Key Achievements

Performance Improvement

1st Run: R² = 0.0045 (Baseline)
2nd Run (v0.1): R² = 0.201 (44x improvement)

Implemented Improvements

✅ Exclude hazard = 0 samples (75% → 0%)
✅ Target normalization with Yeo-Johnson transformation
✅ Feature standardization with StandardScaler
✅ NAM hyperparameter tuning
✅ Ensemble model (FAM + GBAM)
✅ 5-Fold Cross-Validation

Details: Lesson_1st_train.md, Lesson_2nd_train.md

Model List (v0.5)

GLM-GAM: Statistical GAM with splines (pyGAM) - R²=0.195
FAM: Forest Additive Models (Random Forest + OOB) - R²=0.191
GBAM: Gradient Boosting Additive Models (LightGBM) - R²=0.213
TabNAM: TabNet-based Neural Additive Model - R²=0.215 ⭐ Best Individual
AGENTIC ENSEMBLE: AI-Optimized Weighted Average - R²=0.216 🏆 Best Overall
- Optimal Weights: TabNAM 80.5% + GAM 19.5%

Legacy Models (v0.1-v0.3, not in v0.5)

NAM: Neural Additive Models (excluded due to instability)
QRAM: Quantile Regression (excluded due to poor performance)
FOBAM v1: Failed 1-feature RF approach (v0.4, replaced by FAM)

Folder Structure

├── data/                     # Dataset
│   └── 260311v1_inspection_base_and_hazard_estimation.csv # Bridge (v0.5)
├── src/                      # Source code
│   ├── data_preprocessing.py # Preprocessing & feature engineering (v0.5)
│   ├── utils.py              # Common utilities
│   ├── evaluation.py         # Evaluation functions
│   ├── interaction_analysis.py # Damage interaction analysis (v0.2)
│   ├── base_interaction_analysis.py # Base feature interactions (v0.3)
│   ├── agentic_tuner.py      # Agentic Ensemble Weight Tuning (v0.5) ⭐
│   └── models/               # Model implementations
│       ├── glm_gam.py        # Statistical GAM
│       ├── fam.py            # Forest Additive Model (v0.4+)
│       ├── gbam.py           # Gradient Boosting Additive Model
│       ├── nam.py            # Neural Additive Model (v0.4+, TabNAM wrapper)
│       ├── tabnam.py         # TabNAM implementation (v0.8)
│       ├── qram.py           # Quantile Regression (legacy)
│       └── __init__.py
├── 0_NOT_CLASSIFICATION/     # Material Degradation (v0.6)
│   ├── fukuchiyama_bridge_sample_v0.6.csv # Material dataset (v0.6)
│   ├── data_preprocessing_classification.py # Preprocessing (v0.6)
│   ├── run_classification_models_v0.6.py # Model training (v0.6) ⭐
│   ├── Classification_v0.6_Results.md # Results documentation (v0.6) ⭐
│   └── Classification_v0.6_arXiv_Draft.md # Paper draft (v0.6)
├── data_preprocessing_cmapss_v07.py # FD004 preprocessing (v0.7)
├── data_preprocessing_cmapss_v08.py # FD002 preprocessing (v0.8) ⭐
├── run_all_models_v0.7.py    # FD004 training (v0.7)
├── run_all_models_v0.8.py    # FD002 training (v0.8) ⭐
├── visualize_shap_cmapss_v0.7.py # FD004 SHAP (v0.7)
├── visualize_shap_cmapss_v0.8.py # FD002 SHAP (v0.8) ⭐
├── generate_paper_figures_v0.7.py # FD004 figures (v0.7)
├── generate_paper_figures_v0.8.py # FD002 figures (v0.8) ⭐
├── RESULTS_v07_CMAPSS_FD004.md # FD004 documentation (v0.7)
├── RESULTS_v08_CMAPSS.md     # FD002 documentation (v0.8) ⭐
├── COMPARISON_FD002_FD004.md # FD002 vs FD004 comparison (v0.8) ⭐
├── notebooks/                # Jupyter Notebooks
│   └── 02_glm_gam.ipynb
├── outputs/                  # Outputs (models, figures, results)
│   ├── models/               # Saved model artifacts
│   ├── figures/              # All visualizations
│   │   ├── shap/             # SHAP visualizations (v0.1)
│   │   ├── shap_interactions/ # Damage interaction SHAP (v0.2)
│   │   ├── shap_base_interactions/ # Base interaction SHAP (v0.3)
│   │   ├── tabnam_shap_dependence/ # TabNAM SHAP (v0.8) ⭐
│   │   ├── interactions/     # Co-occurrence matrix (v0.2)
│   │   ├── model_comparison_v0.5.png # Paper Figure 1 (v0.5)
│   │   ├── model_comparison_v0.8.png # Paper Figure 1 (v0.8) ⭐
│   │   ├── optimization_process_v0.5.png # Paper Figure 2 (v0.5)
│   │   ├── optimization_process_v0.8.png # Paper Figure 2 (v0.8) ⭐
│   │   ├── optimal_weights_v0.5.png # Paper Figure 3 (v0.5)
│   │   ├── optimal_weights_v0.8.png # Paper Figure 3 (v0.8) ⭐
│   │   ├── ensemble_comparison_v0.5.png # Paper Figure 4 (v0.5)
│   │   ├── ensemble_comparison_v0.8.png # Paper Figure 4 (v0.8) ⭐
│   │   └── summary_table_v0.5.png # Paper Figure 5 (v0.5)
│   │   └── summary_table_v0.8.png # Paper Figure 5 (v0.8) ⭐
│   └── results/
│       ├── all_models_comparison.csv # v0.1 results
│       ├── all_models_comparison_v0.2.csv # v0.2 results
│       ├── all_models_comparison_v0.3.csv # v0.3 results
│       ├── all_models_comparison_v0.5.csv # v0.5 results
│       ├── all_models_comparison_v0.6.csv # v0.6 results (Material) ⭐
│       ├── all_models_comparison_v0.7.csv # v0.7 results (FD004) ⭐
│       ├── all_models_comparison_v0.8.csv # v0.8 results (FD002) ⭐
│       ├── agentic_tuning_v0.5.json # Optimization history (v0.5)
│       ├── agentic_tuning_v0.7.json # Optimization history (v0.7)
│       └── agentic_tuning_v0.8.json # Optimization history (v0.8) ⭐
├── run_all_models.py         # v0.1 main script
├── run_all_models_v0.2.py    # v0.2 main script
├── run_all_models_v0.3.py    # v0.3 main script
├── run_all_models_v0.5.py    # v0.5 main script
├── generate_paper_figures_v0.5.py # Paper visualization script (v0.5)
├── visualize_shap.py         # SHAP visualization (v0.1)
├── visualize_shap_interactions_v0.2.py # Damage interaction viz (v0.2)
├── visualize_shap_base_interactions_v0.3.py # Base interaction viz (v0.3)
├── visualize_shap_tabnam_v0.5.py # TabNAM SHAP (v0.5)
├── Lesson_1st_train.md       # v0.1 lessons learned
├── Lesson_2nd_train.md       # v0.2 lessons learned
├── Lesson_base_interaction.md # v0.3 lessons learned
├── Lesson_damage_interaction.md # v0.2 damage interaction analysis
├── Lesson_4th_train.md       # v0.4 lessons learned (TabNAM, FOBAM)
├── Lesson_5th_train.md       # v0.5 lessons learned (Agentic Tuning)
├── METHODOLOGY.md            # Full academic paper (v0.5)
├── README.md                 # This file
├── RELEASE_NOTES_v0.2.md     # v0.2 release notes
└── requirements.txt          # Dependencies

Usage

1. Environment Setup

pip install -r requirements.txt

2. Run All Models

Bridge Degradation (v0.1-v0.5, Infrastructure Domain)

v0.1 (Base features only - 23 features)

python run_all_models.py

v0.2 (With damage interaction features - 73 features)

python run_all_models_v0.2.py

v0.3 (With base feature interactions - 73 features)

python run_all_models_v0.3.py

v0.5 (With Agentic Ensemble Weight Tuning - 73 features) 🏆

python run_all_models_v0.5.py

Material Degradation (v0.6, Material Science Domain)

v0.6 (Material Accident Prediction - 17 features) 🏆 Best R² = 0.768

python 0_NOT_CLASSIFICATION/run_classification_models_v0.6.py

Jet Engine RUL Prediction (v0.7-v0.8, Aviation Domain)

v0.7 (FD004: 2 Fault Modes - 55 features)

python run_all_models_v0.7.py

v0.8 (FD002: 1 Fault Mode - 55 features) 🏆 Current, 9% improvement

python run_all_models_v0.8.py

Output Files

Bridge (v0.1-v0.5):
- outputs/results/all_models_comparison_v0.5.csv - Best bridge performance ⭐
- outputs/results/agentic_tuning_v0.5.json - Optimization history
- Runtime: ~2-3 minutes (v0.5, includes Agentic Tuning)
Material (v0.6):
- 0_NOT_CLASSIFICATION/classification_models_v0.6.csv - Material degradation results ⭐
- Runtime: ~1-2 minutes
Aviation (v0.7-v0.8):
- outputs/results/all_models_comparison_v0.7.csv - FD004 results
- outputs/results/all_models_comparison_v0.8.csv - FD002 results ⭐
- outputs/results/agentic_tuning_v0.8.json - Optimization history (v0.8)
- Runtime: ~3-4 minutes per version

3. Generate Paper Figures

Bridge (v0.5)

python generate_paper_figures_v0.5.py

Output (300 DPI, publication-ready):

outputs/figures/model_comparison_v0.5.png - R²/RMSE/MAE comparison
outputs/figures/optimization_process_v0.5.png - R² & weight evolution
outputs/figures/optimal_weights_v0.5.png - Final weight distribution
outputs/figures/ensemble_comparison_v0.5.png - 6 model performance
outputs/figures/summary_table_v0.5.png - Performance summary table

Aviation (v0.7, v0.8)

python generate_paper_figures_v0.7.py  # FD004
python generate_paper_figures_v0.8.py  # FD002 ⭐

Output (v0.8):

outputs/figures/model_comparison_v0.8.png - Model comparison
outputs/figures/optimization_process_v0.8.png - Optimization trajectory
outputs/figures/optimal_weights_v0.8.png - Final weights
outputs/figures/ensemble_comparison_v0.8.png - Ensemble performance
outputs/figures/summary_table_v0.8.png - Summary table

4. SHAP Visualization (Interpretability Analysis)

Bridge Models (v0.1-v0.5)

Base features (v0.1)

python visualize_shap.py

Damage interactions (v0.2)

python visualize_shap_interactions_v0.2.py

Base feature interactions (v0.3) ⭐ Recommended for interpretability

python visualize_shap_base_interactions_v0.3.py

TabNAM attention (v0.5)

python visualize_shap_tabnam_v0.5.py

Output:

outputs/figures/shap/ - v0.1 SHAP visualizations
outputs/figures/shap_interactions/ - v0.2 damage interaction analysis
outputs/figures/shap_base_interactions/ - v0.3 base interaction analysis
- interaction_summary.png - SHAP summary for base interactions
- interaction_importance_bar.png - Top 20 base interaction features
- interaction_vs_base_comparison.png - Interaction vs base importance
- interaction_heatmap.png - Feature pair heatmap
- dependence/*.png - Top 8 base interaction dependence plots

Aviation Models (v0.7-v0.8)

FD004 SHAP (v0.7)

python visualize_shap_cmapss_v0.7.py

FD002 SHAP (v0.8) ⭐ Current

python visualize_shap_cmapss_v0.8.py

Output (v0.8):

outputs/figures/tabnam_shap_dependence/ - FD002 SHAP analysis
- 6 SHAP visualization plots
- shap_importances.csv - Feature importance values

5. Detailed Analysis with Jupyter Notebooks

Individual model notebooks in notebooks/:

02_glm_gam.ipynb: GAM model (baseline)

6. Read Documentation

Bridge Degradation (v0.1-v0.5)

Quick Start: QUICKSTART.md
Academic Paper: METHODOLOGY.md - Full methodology & results ⭐
Lessons Learned:
- Lesson_1st_train.md - v0.1 baseline
- Lesson_2nd_train.md - v0.2 damage interactions
- Lesson_base_interaction.md - v0.3 base interactions
- Lesson_damage_interaction.md - v0.2 detailed analysis
- Lesson_4th_train.md - v0.4 TabNAM & FOBAM
- Lesson_5th_train.md - v0.5 Agentic Ensemble ⭐

Material Degradation (v0.6)

Classification_v0.6_Results.md - Full results & analysis ⭐
Classification_v0.6_arXiv_Draft.md - Paper draft

Aviation (v0.7-v0.8)

RESULTS_v07_CMAPSS_FD004.md - FD004 comprehensive results
RESULTS_v08_CMAPSS.md - FD002 comprehensive results ⭐
COMPARISON_FD002_FD004.md - Single vs Dual fault mode comparison ⭐

Target Variable

h_i: Component-damage hazard (1st estimation)
Candidates: hazard_rate_1to2, hazard_rate_2to3, inverse of years_to_rank3

Features

Static Features (x_i) - 15 features

age, construction_year, material_type, bridge_form, bridge_type
bridge_length_m, total_width_m, num_bearings, num_expansion_joints, num_spans
bridge_area_m2, emergency_road, DID_district, bus_route, coastal_zone

Dynamic Features (z_i) - 8 features

inspection_cycle_no, inspection_year
soundness_level, damage_category
damage_progressiveness_score, damage_importance_score
LCC_soundness_II, LCC_soundness_III

Total: 23 features (15 numeric + 8 encoded categorical)

Evaluation Metrics

RMSE (Root Mean Squared Error)
MAE (Mean Absolute Error)
R² (Coefficient of Determination)
Quantile Loss (QRAM only)

Visualization

Additive effect curves for each feature
SHAP values (GBAM)
Partial dependence plots (FAM)
Model performance comparison

Data Splitting

Group K-Fold: Split by bridge_id (prevents data leakage)
5-fold recommended
1,316 unique bridges

Execution Order

GLM-GAM (Baseline & feature effect understanding)
FAM (Feature importance ranking)
NAM (Nonlinearity verification)
GBAM (High-accuracy model)
QRAM (High-risk zone analysis)
ENSEMBLE (Final prediction)

Key Technical Details

Data Preprocessing

Zero hazard filtering: Excluded 88,945 samples (66.4%) with h_i = 0
Target transformation: Yeo-Johnson (λ = -41.04)
Feature scaling: StandardScaler for 15 numeric features
Categorical encoding: Label encoding for 8 categorical features

Cross-Validation

Method: 5-Fold Group K-Fold by bridge_id
Training samples: ~36,046 per fold
Test samples: ~9,012 per fold

Top 5 Important Features (SHAP, GBAM v0.3)

damage_category_encoded (0.2951) - Most important
coastal_zone_encoded (0.0654)
age (0.0255)
construction_year (0.0219)
DID_district_encoded (0.0213)

Top 5 Base Interactions (SHAP, v0.3)

age × LCC_soundness_II (0.0156) - Compound aging deterioration
damage_progressiveness_score × DID_district_binary (0.0136)
LCC_soundness_II × inspection_cycle_no (0.0119)
age × emergency_road_binary (0.0116)
inspection_cycle_no × age (0.0107)

Key Achievements (v0.8 Framework)

Multi-Domain Success (v0.5-v0.8)

🌍 3 Domains: Infrastructure (Bridge), Materials (Accidents), Aviation (Jet Engines)
📊 4 Case Studies: v0.5 (R²=0.216), v0.6 (R²=0.768), v0.7 (R²=0.609), v0.8 (R²=0.663)
🏆 Best Performance: Material degradation (v0.6, R²=0.768) - highest across all studies
🎯 Key Finding: Single fault mode (FD002) outperforms dual fault mode (FD004) by 9%

Performance by Domain

Domain	Case Study	R²	Dataset Size	Complexity
Material	v0.6	0.768	3.8K	Low (specific patterns)
Aviation	v0.8 (FD002)	0.663	54K	Medium (1 fault)
Aviation	v0.7 (FD004)	0.609	61K	High (2 faults)
Infrastructure	v0.5 (Bridge)	0.216	45K	Very High (diverse damage)

Innovation (v0.5)

World's First: Agentic AI for ensemble weight optimization in GAMs
Planner-Executor-Evaluator: 3-agent architecture for intelligent search
Fast Convergence: Optimal solution found in iteration 1 (of 12)
TabNAM: First application of TabNet attention to Neural Additive Models

Domain-Specific Insights

Material Degradation (v0.6):

TabNAM dominates (64.8% weight) - effective for focused patterns
Small dataset (3.8K) achieves highest R² - domain specificity matters
Clear material science deterioration patterns

Jet Engine FD002 (v0.8):

Single fault mode → 9% higher R² than FD004
3-model ensemble (GAM + GBAM + TabNAM) vs. 2-model in FD004
TabNAM returns (22.1% weight) after being 0% in FD004
All individual models perform better than FD004 counterparts

Jet Engine FD004 (v0.7):

Dual fault modes (HPC + Fan) increase complexity
2-model ensemble (GAM + GBAM only) - TabNAM ineffective
Temperature sensors dominate feature importance

Bridge Degradation (v0.5):

Most complex problem (32 damage types, diverse infrastructure)
44x improvement from baseline (R² = 0.0045 → 0.216)
Agentic Ensemble: TabNAM 80.5% + GAM 19.5% optimal

Interpretability

Additive Structure: All models preserve f(x) = Σ f_i(x_i) interpretability
SHAP Analysis: Comprehensive feature importance across all domains
Engineering-Meaningful: Domain-specific insights from additive components
Cross-Domain: Framework adapts model selection to problem complexity

Production-Ready

Diverse Scales: 3.8K to 260K samples across case studies
5-Fold CV: Robust group-based validation
Computational Efficiency: 1-4 minutes per version
Paper-Ready: Publication-quality figures for all versions

Requirements

Python 3.12+
PyTorch 2.5.1
LightGBM 4.5.0
SHAP 0.46.0
pygam 0.9.1
scikit-learn 1.6.1
pandas 2.2.3
numpy 2.2.3
pytorch-tabnet 4.0 (v0.4+)
scipy 1.14.1 (v0.5+)

See requirements.txt for complete list.

License

MIT License

Citation

If you use this work, please cite:

@software{diverse_gams_v0.8,
  title={Diverse GAMs: Multi-Domain Degradation Prediction Framework},
  author={[Your Name]},
  year={2026},
  version={0.8},
  note={Agentic Ensemble and TabNAM for Infrastructure, Materials, and Aviation}
}

Case Study Citations:

v0.5 (Bridge): Agentic Ensemble Weight Tuning for Bridge Hazard Prediction
v0.6 (Material): Material Degradation Accident Prediction with TabNAM
v0.7 (Aviation FD004): Dual Fault Mode Jet Engine RUL Prediction
v0.8 (Aviation FD002): Single Fault Mode Jet Engine RUL Prediction (9% improvement)

Acknowledgments

pyGAM: Statistical GAM implementation
LightGBM: High-performance gradient boosting
SHAP: Model interpretability framework
TabNet: Attention-based tabular learning (pytorch-tabnet)
NASA C-MAPSS: Turbofan Engine Degradation Simulation Dataset (v0.7-v0.8)

Contact

For questions or collaboration: [Your Contact Information]

Project Status: ✅ 4 Case Studies Complete (v0.8 Current) Last Updated: March 15, 2026 Repository: [Your Repository URL]

Name		Name	Last commit message	Last commit date
Latest commit History 7 Commits
data		data
notebooks		notebooks
outputs		outputs
src		src
.gitignore		.gitignore
COMPARISON_3Cases.md		COMPARISON_3Cases.md
COMPARISON_FD002_FD004.md		COMPARISON_FD002_FD004.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
Lesson_5th_train.md		Lesson_5th_train.md
METHODOLOGY.md		METHODOLOGY.md
QUICKSTART.md		QUICKSTART.md
README.md		README.md
RESULTS_v06_Material.md		RESULTS_v06_Material.md
RESULTS_v07_CMAPSS.md		RESULTS_v07_CMAPSS.md
RESULTS_v08_CMAPSS.md		RESULTS_v08_CMAPSS.md
data_preprocessing_cmapss_v07.py		data_preprocessing_cmapss_v07.py
data_preprocessing_cmapss_v08.py		data_preprocessing_cmapss_v08.py
data_preprocessing_material_v06.py		data_preprocessing_material_v06.py
explore_cmapss_data_v0.7.py		explore_cmapss_data_v0.7.py
explore_material_data_v0.6.py		explore_material_data_v0.6.py
generate_paper_figures_v0.5.py		generate_paper_figures_v0.5.py
generate_paper_figures_v0.6.py		generate_paper_figures_v0.6.py
generate_paper_figures_v0.7.py		generate_paper_figures_v0.7.py
generate_paper_figures_v0.8.py		generate_paper_figures_v0.8.py
inspect_material_data.py		inspect_material_data.py
requirements.txt		requirements.txt
run_all_models_v0.5.py		run_all_models_v0.5.py
run_all_models_v0.6.py		run_all_models_v0.6.py
run_all_models_v0.7.py		run_all_models_v0.7.py
run_all_models_v0.8.py		run_all_models_v0.8.py
visualize_shap_cmapss_v0.7.py		visualize_shap_cmapss_v0.7.py
visualize_shap_cmapss_v0.8.py		visualize_shap_cmapss_v0.8.py
visualize_shap_material_v0.6.py		visualize_shap_material_v0.6.py
visualize_shap_tabnam_v0.5.py		visualize_shap_tabnam_v0.5.py

Folders and files

Latest commit

History

Repository files navigation

Diverse Additive Models: Multi-Domain Degradation Prediction Framework

Description

Version History

Bridge Degradation (Infrastructure Domain)

Multi-Domain Expansion

What's New in v0.8 (Current) 🎉

Jet Engine RUL Prediction - Single Fault Mode (FD002)

Why FD002 > FD004?

Top Features (SHAP, FD002)

What's New in v0.7 🚁

Jet Engine RUL Prediction - Dual Fault Modes (FD004)

Key Findings

Top Features (SHAP, FD004)

What's New in v0.6 🏭

Material Degradation Accident Prediction

Key Findings

Top Features (v0.6)

What's New in v0.5 🎉

Agentic Ensemble Weight Tuning

Key Findings

TabNAM (v0.4-v0.5)

What's New in v0.4

Diverse f(·) Exploration

FOBAM v2 Redesign

What's New in v0.3

Base Feature Interaction Analysis

Project Overview

Multi-Domain Framework (v0.8)

Key Insights

Original Bridge Study (v0.1-v0.5)

v0.1 Key Achievements

Performance Improvement

Implemented Improvements

Model List (v0.5)

Legacy Models (v0.1-v0.3, not in v0.5)

Folder Structure

Usage

1. Environment Setup

2. Run All Models

Bridge Degradation (v0.1-v0.5, Infrastructure Domain)

Material Degradation (v0.6, Material Science Domain)

Jet Engine RUL Prediction (v0.7-v0.8, Aviation Domain)

Output Files

3. Generate Paper Figures

Bridge (v0.5)

Aviation (v0.7, v0.8)

4. SHAP Visualization (Interpretability Analysis)

Bridge Models (v0.1-v0.5)

Aviation Models (v0.7-v0.8)

5. Detailed Analysis with Jupyter Notebooks

6. Read Documentation

Bridge Degradation (v0.1-v0.5)

Material Degradation (v0.6)

Aviation (v0.7-v0.8)

Target Variable

Features

Static Features (x_i) - 15 features

Dynamic Features (z_i) - 8 features

Evaluation Metrics

Visualization

Data Splitting

Execution Order

Key Technical Details

Data Preprocessing

Cross-Validation

Top 5 Important Features (SHAP, GBAM v0.3)

Top 5 Base Interactions (SHAP, v0.3)

Key Achievements (v0.8 Framework)

Multi-Domain Success (v0.5-v0.8)

Performance by Domain

Innovation (v0.5)

Domain-Specific Insights

Interpretability

Production-Ready

Requirements

License

Packages