↓ Skip to main content
blanpa
  1. Projects/

Condition Monitoring Suite

Node-RED JavaScript Python Machine Learning IIoT
Table of Contents

The Problem
#

Unplanned machine downtime costs manufacturing companies thousands of euros per hour. Traditional maintenance approaches — either reactive (fix when broken) or calendar-based (replace after X months) — are either too late or too wasteful. Sensors generate massive amounts of data, but without the right tools, operators can’t act on it in time.

The Solution
#

A Node-RED module that turns raw sensor data into actionable maintenance insights — directly on the factory floor, without cloud dependency. 9 specialized nodes cover the full predictive maintenance pipeline: from data ingestion and anomaly detection to ML inference and health scoring.

GitHub Repository Thumbnail
blanpa/node-red-contrib-condition-monitoring

node-red-contrib-condition-monitoring

Jupyter Notebook
1
1

Architecture
#

The suite is designed as a modular pipeline where each node handles one responsibility:

Sensors → Signal Analyzer → Anomaly Detector → Health Index → Dashboard/Alert
                ↓                   ↓
        Training Data          Trend Predictor
          Collector               (RUL)

Anomaly Detection
#

10 statistical algorithms to catch deviations before they become failures:

AlgorithmBest For
Z-ScoreNormally distributed sensor data
IQROutlier detection with skewed distributions
CUSUMDetecting small, gradual shifts
Isolation ForestUnsupervised ML, no training labels needed
PCAMultivariate anomaly detection across correlated sensors

All detectors support hysteresis — preventing alarm floods from noisy signals oscillating near thresholds.

Signal Processing
#

  • FFT (Radix-4 Cooley-Tukey) for frequency analysis of vibration data
  • Envelope detection for bearing fault diagnosis
  • Cepstrum analysis for gearbox diagnostics
  • ISO 10816-3 vibration severity classification (zones A-D)
  • Butterworth filtering with zero-phase processing

Predictive Maintenance
#

  • Remaining Useful Life (RUL) — Weibull distribution modeling predicts when a component will likely fail
  • Trend prediction — Linear regression and exponential smoothing forecast sensor trajectories
  • Health Index — Multi-sensor aggregation with dynamic weighting produces a single 0-100 health score

ML Inference at the Edge
#

Run trained models directly in Node-RED without cloud roundtrips:

  • ONNX, TensorFlow.js, Keras, scikit-learn, TFLite
  • Google Coral / Edge TPU hardware acceleration for real-time inference
  • Persistent Python subprocess bridge for efficient repeated inference
  • Built-in Training Data Collector exports labeled datasets to CSV/JSONL for model retraining

Quality
#

  • 135 unit tests covering all nodes and edge cases
  • State persistence across Node-RED restarts
  • Dynamic runtime configuration via message objects
  • MIT licensed, production-ready

Related