Predictive Maintenance for Power Turbines
On-Device Continual Learning (TinyML)
The Problem
A hydro-electric plant needed to monitor turbine bearings. Traditional vibration analysis only detected failure after significant damage occurred. Because turbine signatures change based on water flow and age, a static model became obsolete every few months.
Solution
Klyff implemented an Unsupervised Anomaly Detection model that utilizes “Continual Learning.” The device creates a “Golden Fingerprint” of the specific turbine’s acoustic and vibration profile and adapts that baseline in real-time as the machinery naturally wears.
Hardware
- MCU: STMicroelectronics STM32N6 (with integrated NPU).
- Sensors: High-frequency 3-axis industrial Accelerometers + Ultrasonic microphones.
Software
Inference Engine: NanoEdge AI Studio integration for on-device training.
Protocol: MQTT-SN over LoRaWAN for ultra-low power status reporting to the Klyff Dashboard.
Results
- Federated Learning learns continuously and updates the fingerprint
- Lead Time: System now provides 22 days of advanced warning before mechanical failure.
- Operational Savings: Avoided a catastrophic failure in Q3, saving the utility $450,000 in emergency repair costs.
Solution Flow
Klyff gives a significant advantage over traditional Cloud IoT
Most IoT platforms are just “pipes” for data. Klyff is the intelligence at the end of the pipe. We focus on the Silicon-to-Model optimization unlike Cloud Dashboards
| Feature | Traditional Cloud IoT | Klyff "Hard Tech" (Edge-First) | The Klyff Advantage |
|---|---|---|---|
| Response Latency | 200ms – 2 seconds (Cloud Round-trip) | < 10ms – 20ms (On-Device Inference) | 10-50x Faster. Essential for high-speed safety stops and robotics. |
| Data Transmission Costs | High. Constant streaming of raw sensor/video data. | Near-Zero. Only anomaly "events" are sent to the cloud. | 90% Cost Reduction. Saves thousands in cloud egress and storage fees. |
| Data Privacy & Security | High Risk. Raw sensitive data travels over the public internet. | Secure by Design. Raw data never leaves the device; only insights move. | GDPR/Union Compliant. Ideal for worker monitoring and IP-sensitive zones. |
| Connectivity Reliance | System fails or "freezes" if Wi-Fi/4G drops. | 100% Autonomous. Decisions continue even during network outages. | Critical Reliability. No downtime for remote mines or offshore assets. |
| Hardware Efficiency | Heavy data logging drains battery and bandwidth. | Hardware-Optimized. Uses NPUs/FPGAs to run AI at 10,000x higher efficiency. | Extended Asset Life. Models run on ultra-low-power $5–$20 chips. |
| Failure Detection | Reactive. Alerts you after a threshold is crossed. | Proactive. Identifies "micro-drifts" in machine behavior weeks earlier. | True Predictive Maintenance. Moves from "alerting" to "preventing." |
Other Case Studies
Selected Customer Success Stories
Adaptive Predictive Maintenance
Solder Joint Inspection
