Our team is building a predictive maintenance system for turbines. We are seeing high latency when sending raw high-frequency vibration data to AWS. Is it better to run the ML models directly on the edge gateway, or should we optimize our data compression before sending it to the cloud?
3 answers
For high-frequency vibration data, Edge Computing is almost always the better choice. Sending raw waveforms to the cloud is expensive in terms of bandwidth and adds too much latency for emergency shutdowns. I recommend using "Edge Intelligence" where you run a lightweight TensorFlow Lite model on the gateway to detect anomalies in real-time. You only send the summary data or the "events" to the cloud for long-term trend analysis and model retraining. This "Fog Computing" approach saves on egress costs and ensures that if your internet connection drops, the turbine’s safety logic still functions locally.
How much local storage does your current gateway have, and can it handle the heat generated by running complex inference models 24/7 in an industrial setting?
A hybrid approach is best. Use the edge for the "Stop" command and the cloud for the "Deep Learning" that improves the model over time.
Perfectly stated, Michael. The cloud should be the 'brain' that learns, while the edge is the 'reflex' that acts immediately on the ground.
That was a major concern, Charles. We ended up upgrading to fanless industrial PCs with high-end thermal dissipation. We found that standard gateways were throttling the CPU during peak vibration analysis. By moving to more ruggedized hardware, we successfully reduced our cloud data footprint by 90% while improving our response time from 2 seconds to under 50 milliseconds. The hardware cost was high, but the savings on our monthly AWS bill have already paid for the upgrade.