Process quality online monitoring + PHM equipment health & fault diagnosis — neural-network-driven device-side sensing (CNN/TCN/VAE fusion), moving from reactive repair to proactive prevention. A pioneer of non-vision AI signal diagnosis in manufacturing
32ch
Synchronous vibration / temperature / RPM acquisition
31K
1D-CNN lightweight model parameters
0.6ms
Edge-side inference per window
3×
CNN + TCN + VAE model fusion
RK3588
ARM64 edge deployment
Process Quality Online Monitoring System
Online monitoring for ultrasonic welding, rotating-part vibration and more — CNN+VAE+rule engine three-path fusion for verdicts; multi-channel real-time acquisition of power, amplitude and frequency, evolving in three stages (diagnosis → prediction → closed loop) and already deployed at scale on welding lines.
🧠
CNN
1D-CNN weld quality classification with a dedicated encoder
🔍
VAE
Variational autoencoder anomaly detection — two paths that complement each other
📐
Rule Engine
Power/amplitude/frequency drift detection with explainable verdicts
🔗
Three-Path Fusion
CNN + VAE + rule engine evolving in a three-stage closed loop
📡
Multi-Channel Acquisition
Synchronous monitoring of power, amplitude and frequency
🏭
Line Deployment
Already running at scale on welding lines
Why three-path fusion? CNN learns known defect patterns from waveforms, VAE discovers anomalies never seen before, and the rule engine quantifies power, amplitude and frequency drift into explainable process criteria — three paths that cross-validate each other, keeping both AI recognition power and verdicts engineers can understand. Defective parts are intercepted instantly, weld data is fully retained, and quality issues stay traceable.
CNNVAERule EngineOnline Monitoring
PHM Equipment Health & Fault Diagnosis System
Equipment health management (PHM) — multi-source sensing of vibration, temperature and current, AI fault diagnosis, and predictive maintenance that cuts unplanned downtime.
📡
Multi-Source Sensing
32-channel synchronous acquisition of vibration, temperature and RPM signals
Dilated-convolution remaining-useful-life prediction that replaces LSTM
🔍
VAE Anomaly Detection
Variational autoencoder that surfaces unknown fault modes
🌲
Isolation Forest
Fast anomaly screening on statistical features to assist verdicts
⚡
RK3588 Edge Unit
Parallel ARM64 inference, deployed at scale in wind farms
Three models, three jobs: 1D-CNN identifies known fault types (bearings, gears and other typical patterns), TCN predicts remaining useful life while the machine is still healthy, and VAE catches new anomalies never seen in the training set — each problem class handled by the right tool. At just 31K total parameters, all three run in real-time parallel on edge hardware like the RK3588, with data never leaving the plant.
PHM1D-CNNTCNVAEPredictive Maintenance
Use Cases
Typical deployments for device-side intelligent sensing
🌬️
Wind Farms
PHM deployed at scale: 32-channel synchronous acquisition of drivetrain vibration, temperature and RPM, with three-model fusion for predictive maintenance — lower unplanned downtime and operations costs.
🔩
Ultrasonic Welding
Real-time power/amplitude/frequency acquisition with CNN+VAE+rule engine fusion to classify good vs. defective welds — defects intercepted instantly, quality fully traceable.
⚙️
Rotating Machinery
Health monitoring for motors, pumps and fans: typical bearing/gear fault pattern recognition and remaining-useful-life prediction — from reactive repair to proactive prevention.
FAQ
Does integration require modifying our equipment?
No changes to the machine itself. Adding an edge computing controller plus vibration/temperature/RPM sensors is enough to capture signals; inference runs at the edge, independent of the cloud, and the machine's existing control system is untouched.
How is data security ensured?
All model inference runs locally on the edge unit (RK3588); raw vibration data never leaves the plant. When cloud reporting is needed, only diagnosis results and statistical features are uploaded, over TLS-encrypted channels.
How soon can the system go live?
A standard deployment is commissioned in 2–4 weeks. If a process model must be trained for a specific machine type, model tuning starts once real production data accumulates — existing wind farm and welding line deployments give us a head start.
Can it connect to our existing MES / management platform?
Yes. Diagnosis results are exported via MQTT/HTTP and can integrate with the ONLINE Data Management & Analytics Platform or your existing MES, with alerts pushed at the same time.
Related Products
Device-side data flows upward, working across the full TSingley cloud-network-edge-device stack