EdgePredict-CWRU

Embedded predictive maintenance system for bearing fault diagnosis, running fully on a Raspberry Pi 4 without any Cloud dependency.

Overview

This project implements an Edge AI pipeline that classifies bearing faults in real time using the CWRU Bearing Dataset.

The system runs entirely on-device:

• Signal ingestion and overlapping windowing
• Statistical feature extraction (RMS, Crest Factor, Kurtosis, Skewness)
• Machine Learning inference with a tuned Random Forest classifier
• Web supervision dashboard served by Flask

Architecture

Three concurrent modules run on the Raspberry Pi:

• Module A: Signal ingestion and feature extraction
• Module B: On-device inference engine
• Module C: Flask web server and Chart.js dashboard

Hardware

• Raspberry Pi 4 Model B
• Cortex-A72, four cores
• 3.7 GiB RAM
• 8 GB USB drive used as boot and storage device (no microSD)

Software stack

• Python 3.13
• NumPy, SciPy
• Scikit-Learn, Joblib
• Flask
• Chart.js
• systemd

Dataset

CWRU Bearing Dataset, 12k Drive End subset.

• 48 files (4 unavailable from public mirrors)
• 4 classes: Normal, InnerRace, OuterRace, Ball
• Sampling frequency: 12 kHz
• Window size: 1024 points
• Step: 128 points (87.5 percent overlap)
• 54781 windows generated

Class distribution:

• InnerRace: 15100
• Ball: 15095
• Normal: 13231
• OuterRace: 11356

Machine Learning pipeline

Six algorithms were compared under a frozen protocol:

Model	CV accuracy	Test accuracy	Inference (ms/sample)	Size (KB)
RandomForest	0.9807 +/- 0.0014	0.9792	0.0133	15456.4
GradientBoosting	0.9783 +/- 0.0016	0.9760	0.0113	513.4
MLP	0.9780 +/- 0.0014	0.9759	0.0033	91.2
DecisionTree	0.9721 +/- 0.0012	0.9699	0.0003	171.6
KNN	0.9608 +/- 0.0012	0.9617	0.0111	3617.4
SVM_RBF	0.9512 +/- 0.0034	0.9586	1.0810	829.0

Random Forest was selected for its best accuracy, best CV mean, lowest CV variance, and fast inference.

Hyperparameter tuning with GridSearchCV (36 combinations, 3-fold CV):

• n_estimators: 50
• max_depth: 20
• min_samples_split: 2
• min_samples_leaf: 1

Final tuned model test accuracy: 0.9779.

Feature importance:

• RMS: 0.6512
• Kurtosis: 0.1843
• Skewness: 0.0951
• Crest Factor: 0.0694

Inference performance

Latency per window: 0.0133 ms per sample in batch inference.

This is far below the 10 ms target from the specification.

Runtime footprint

Measured with htop at steady state:

• CPU usage: 4 to 9 percent
• Memory: 162 MB for the Python process
• Load average: 0.75 / 0.64 / 0.45 on a 4-core system

Deployment

The application runs as a systemd service:

sudo systemctl status edgepredictpi.service

Dashboard access:

http://<raspberry-pi-ip>:5000

Documentation

• docs/dataset.md
• docs/features.md
• docs/model.md
• docs/model_comparison.md
• docs/tuning.md
• docs/final_evaluation.md
• docs/ia_summary.md
• docs/performance.md
• docs/api.md
• docs/deployment.md
• docs/test_plan.md

Report

The technical report is available at report/report.md.

Notes

• The GridSearchCV tuning was performed on the development machine because the Raspberry Pi 4 without active cooling reached thermal throttling at 84.7 degrees Celsius under sustained multi-core load.
• The tuned model was then transferred to the Pi for deployment.
• This reflects a standard MLOps workflow: train and tune off-device, deploy the final artifact on the target.

Author

Horacia Azonhoumon GSEA - Embedded Electronics and Automation ENSA Tanger

Report and demonstration

• Technical report: report/EdgePredict-CWRU-report.pdf
• Demonstration screenshots: docs/demo/
• Video demonstration: [link]

Repository structure

src/                  Python source files
templates/            Dashboard HTML
static/               Static assets
systemd/              Systemd service unit
docs/                 Documentation and demo
report/               LaTeX sources and final PDF
data/                 Not versioned (raw dataset and features)
models/               Not versioned (trained model)