Resilient Edge: Architectural Optimization for Real-Time Network Fault Diagnosis on Resource-Constrained Embedded Systems

Experimental Dataset & Codebase: TinyML for Embedded Network Fault Diagnosis This repository contains the complete curated training datasets, Python machine learning pipeline, and C++ hardware deployment code from a longitudinal experimental campaign evaluating the "Resilient Edge" TinyML framework on severely resource-constrained ESP32 microcontrollers. The codebase and data were generated to validate a novel Decision Tree Distillation technique and physics-based feature extraction pipeline designed for noisy and attenuated 2.4 GHz wireless environments. Note: The standalone, comprehensive raw telemetry dataset corresponding to this work is available at DOI: 10.5281/zenodo.18470802. 📄 ASSOCIATED MANUSCRIPT This dataset and codebase support the findings presented in the research article: "Resilient Edge: Architectural Optimization for Real-Time Network Fault Diagnosis on Resource-Constrained Embedded Systems" 📂 REPOSITORY STRUCTURE Resilient_Edge_TinyML_Codebase/ │ ├── 01_Raw_Data/ (Original Server Logs - ESP32 to Gateway) ├── 02_Processed_Data/ (Curated CSVs after ETL processing) ├── 03_Models/ (Compiled .pkl binaries & .h C-headers) ├── 04_ESP32_Hardware_Code/ (Arduino sketch for TinyML deployment) │ └── TinyML-Experiment/ │ ├── TinyML-Experiment.ino │ └── model.h ├── 05_Python_Codebase/ (Modular ML Pipeline & Reproducibility) │ ├── 01_ETL_Pipeline_load_inspect.py │ ├── ... │ └── 15_Energy_Comparison.py ├── 06_Exported_Figures/ (High-resolution manuscript plots) └── README.md 📊 KEY DATA / COLUMN DEFINITIONS R_i: Extracted resistance/congestion feature from the wireless channel. L_i: Extracted inductance/attenuation feature. J_i: Extracted jitter/jerk feature from packet dynamics. latency_us: On-device inference latency measured in microseconds (µs). fault_class: Ground truth label for the diagnosed network state. 🛠 USAGE The included Python scripts in the 05_Python_Codebase/ folder are numbered sequentially (01 to 15). Running these scripts in order will process the datasets, train the AI models, and automatically reproduce the Speed vs. Accuracy Pareto optimization figures found in the associated manuscript. To replicate the hardware benchmarks (e.g., the 0.26 µs inference latency), compile and flash the TinyML-Experiment.ino file directly to an ESP32 microcontroller. Environment: Python 3.x, Scikit-Learn, Pandas, Matplotlib, Seaborn, Arduino IDE (v1.8+ or v2.x with ESP32 Core).