Home / Projects / atlas-copco-iiot

IIoT Predictive Maintenance — Atlas Copco G15 FF

Edge ML pipeline for non-invasive anomaly detection on industrial air compressors using contact-acoustic sensors, DSP feature extraction, and a CNN fault classifier.

PythonPyTorchNumPySciPyMQTTInfluxDBGrafanaRaspberry Pi 4Maijker Magnetic Sensors

Problem Statement

Atlas Copco G15 FF rotary screw compressors are workhorses in industrial facilities, but unplanned failures cost thousands in lost production per event. Traditional vibration monitoring requires OEM-installed accelerometers and proprietary software. We needed a non-invasive retrofit: attach a sensor without disassembly, stream telemetry to the cloud, and classify fault modes in near-real-time — all from commodity hardware.

Technical Details

Part of Purdue ME 59700 (Industrial IoT), this project implements a full-stack monitoring pipeline for the Atlas Copco G15 FF air compressor — from raw acoustic signal to cloud dashboard.

Architecture

  1. Sensing: Maijker magnetic contact-acoustic sensors clamp directly to the compressor housing — no disassembly, no drilling. They capture structure-borne vibration up to ~20 kHz.
  2. Edge Processing: Raw signal sampled at 10 kHz on a Raspberry Pi 4. Windowed FFT → feature vector: RMS energy, kurtosis, spectral centroid, per-band power in four sub-bands.
  3. Inference: Lightweight 1D-CNN (≈40k parameters) trained on labeled fault signatures. Three output classes: baseline, bearing wear, valve degradation. Inference runs in <50ms per window.
  4. Telemetry: Inference results + raw features published over MQTT to a cloud broker. Time-series stored in InfluxDB; Grafana dashboard for live monitoring and alerting.

Model Training

Training data collected during controlled fault-injection sessions on a test unit. Bearing wear simulated via seeded surface defects; valve degradation via partial obstruction. 80/20 train-test split, stratified by fault class and session.

Current Results

Early prototype achieves >92% classification accuracy on held-out test data across all three fault classes. Ongoing work: INT8 quantization for Cortex-M deployment, expanding fault catalog to include motor overheating signatures.

Lessons Learned

Non-invasive sensing is harder than it sounds — sensor placement and coupling pressure dramatically affect signal quality. Spent more time on sensor mounting fixtures than on the ML. Also: always validate your FFT windowing before training; a subtle off-by-one in the hop size caused two days of mysteriously poor validation loss.