The default architecture for IoT bearing monitoring has been cloud-centric: sensors capture vibration data, transmit it to a cloud platform, and algorithms running on remote servers perform analysis, anomaly detection, and alerting. This approach works — until it doesn’t. Latency, bandwidth costs, connectivity failures, and data sovereignty concerns are driving a shift toward edge AI, where anomaly detection and bearing defect classification run locally on the gateway rather than in the cloud. This article examines the architectural trade-offs and explains when edge processing is the right choice for bearing condition monitoring.
The Cloud-Centric Model and Its Limitations
In a typical cloud-based bearing monitoring architecture, sensors capture vibration data and transmit it wirelessly to a gateway. The gateway forwards the data (often via cellular or satellite) to a cloud platform — AWS IoT Core, Azure IoT Hub, or a vendor-specific SaaS platform. Cloud servers run the analytics: FFT computation, bearing defect frequency identification, anomaly detection models, and health scoring. Results are pushed back to dashboards and alerting systems.
This model has clear advantages: elastic compute resources, centralized model management, easy software updates, and the ability to aggregate data across many sites for fleet-level analytics. But it has equally clear failure modes.
Latency
The round-trip time from sensor to cloud to alert can range from seconds to minutes, depending on connectivity, queue depths, and processing load. For most bearing monitoring applications, this latency is acceptable — bearing degradation unfolds over days or weeks, not seconds. But for applications where rapid response matters (e.g., automatic shutdowns, safety interlocks, or real-time operator alerts during commissioning), cloud latency introduces unacceptable delay.
Bandwidth and Cost
Transmitting raw or semi-processed vibration data over cellular networks is expensive. A single sensor capturing 2-second windows at 25.6 kHz every 15 minutes generates approximately 3.4 MB/hour of raw data. For a facility with 50 sensors, that’s 170 MB/hour or roughly 4 GB/day. At typical industrial IoT cellular rates of $1–5/GB, the data transmission cost alone can exceed $120–600/month — often more than the hardware amortization cost.
Edge processing reduces this dramatically. If anomaly detection and feature extraction happen on the gateway, only compact summary data and alerts need to traverse the cellular link. The same 50-sensor deployment might transmit 50–200 KB/hour instead of 170 MB/hour — a reduction of 99.9%, bringing monthly data costs below $5.
Connectivity Dependency
Cloud-dependent systems fail when connectivity fails. This is not a theoretical concern — it’s a daily reality in many industrial environments:
- Marine vessels lose cellular connectivity when more than 20–50 km offshore. Satellite links (Iridium, Starlink) provide intermittent, expensive, and bandwidth-limited alternatives.
- Mining operations in remote locations may have unreliable cellular coverage or operate in radio-quiet zones.
- Railway rolling stock passes through tunnels, rural dead zones, and areas with congested networks.
- Industrial facilities in developing regions may have unstable internet infrastructure.
An edge AI system continues to monitor, detect anomalies, and generate alerts regardless of connectivity status. Data synchronizes to the cloud when connectivity is available, but the monitoring function is never interrupted.
Data Sovereignty and Security
Vibration data from industrial equipment is more sensitive than many organizations realize. High-resolution vibration signatures can reveal:
- Production rates and machine utilization
- Process parameters and operating conditions
- Equipment age and condition (competitive intelligence)
- Maintenance practices and compliance status
For defense contractors, energy infrastructure, pharmaceutical manufacturing, and other regulated or sensitive industries, transmitting this data to third-party cloud platforms raises legitimate security and compliance concerns. Edge processing keeps the raw data on-premises, transmitting only aggregated metrics and alerts.
What Edge AI Actually Does for Bearing Monitoring
The term “edge AI” covers a spectrum of capabilities. For bearing condition monitoring, the relevant functions include:
1. Anomaly Detection
The most fundamental edge AI function is answering the question: is this vibration pattern normal or abnormal? Anomaly detection models learn the baseline vibration signature of each monitored bearing during a training period, then flag statistical deviations from that baseline.
Common approaches include:
- Statistical methods: Z-score monitoring of RMS, peak, and crest factor values against historical distributions
- Autoencoder neural networks: Trained on normal vibration patterns, these models produce high reconstruction error when presented with anomalous data
- Isolation forests: Ensemble methods that efficiently identify outliers in multi-dimensional feature spaces
These models are lightweight enough to run on gateway-class hardware (ARM Cortex-A processors, 1–4 GB RAM) with inference times measured in milliseconds.
2. Bearing Defect Classification
Beyond detecting that something is wrong, edge AI can classify what is wrong. Bearing defects produce characteristic frequency signatures:
- BPFO (Ball Pass Frequency Outer Race): Outer race defect — typically the most common bearing failure mode
- BPFI (Ball Pass Frequency Inner Race): Inner race defect
- BSF (Ball Spin Frequency): Rolling element defect
- FTF (Fundamental Train Frequency): Cage defect
A classification model trained on labeled spectral data can identify which defect type is developing and estimate its severity — enabling maintenance teams to prioritize interventions and order the correct replacement parts before the failure occurs.
3. Health Scoring
Edge AI can compute a composite health score for each bearing, integrating multiple indicators (RMS trend, spectral energy in defect frequency bands, temperature, crest factor trend) into a single 0–100 score. This abstraction makes the data accessible to operators and maintenance planners who may not be vibration analysis specialists.
Health score algorithms can be as simple as weighted threshold models or as sophisticated as gradient-boosted regression models trained on historical failure data. The key is that the computation happens locally, and the score is available immediately — no cloud round-trip required.
4. Adaptive Sampling
An underappreciated edge AI function is adaptive sampling — dynamically adjusting the sensor’s measurement frequency based on detected conditions. When a bearing is healthy, the sensor can sample infrequently (every 30–60 minutes) to conserve battery. When the anomaly detection model detects a developing fault, it signals the sensor to increase sampling frequency (every 1–5 minutes) for higher temporal resolution during the critical degradation period.
This feedback loop between edge intelligence and sensor behavior is only possible when the AI runs locally. A cloud-based system introduces too much latency to make real-time sampling decisions.
Edge vs. Cloud vs. Hybrid: Architecture Comparison
| Criterion | Cloud-Only | Edge-Only | Hybrid (Edge + Cloud) |
|---|---|---|---|
| Alert latency | Seconds to minutes | Milliseconds | Milliseconds (edge) + enriched alerts (cloud) |
| Connectivity required | Always | Never | For sync only — monitoring continues offline |
| Data transmission cost | High ($100+/mo for 50 sensors) | Near zero | Low ($5–20/mo for 50 sensors) |
| Model update complexity | Simple (server-side deploy) | Moderate (OTA gateway update) | Both available |
| Fleet-level analytics | Native | Not available | Available when synced |
| Data sovereignty | Data leaves premises | Data stays on-premises | Raw data stays; summaries sync |
| Hardware requirements | Minimal gateway | Capable gateway (ARM + RAM) | Capable gateway |
For most industrial bearing monitoring deployments, the hybrid architecture offers the best balance. Edge AI handles real-time monitoring, anomaly detection, and alerting. Cloud services provide fleet-level analytics, long-term trend storage, model training on aggregated data, and remote management. The edge functions independently when connectivity is unavailable; the cloud enriches the analysis when connectivity is present.
Gateway Hardware for Edge AI
Running AI models at the edge requires more capable gateway hardware than a simple data forwarder. Typical requirements include:
- Processor: ARM Cortex-A53/A72 or equivalent (e.g., Raspberry Pi Compute Module, NVIDIA Jetson Nano, NXP i.MX8)
- RAM: 1–4 GB for model inference and data buffering
- Storage: 16–64 GB for local data retention and model storage
- Connectivity: BLE and/or LoRa radio for sensor communication; Ethernet, Wi-Fi, or cellular for cloud sync
- Power: 5–15W continuous — typically mains-powered, though battery-backed options exist for remote sites
The inference workload for bearing condition monitoring is modest by modern AI standards. A classification model running on a Cortex-A53 can process FFT features from 50+ sensors in under a second. This is not large language models or computer vision — it’s compact, specialized signal processing models operating on structured numerical data.
When Edge AI Matters Most
Edge AI is not universally necessary for every bearing monitoring deployment. It provides the greatest value in environments where:
- Connectivity is unreliable or expensive: Marine, mining, railway, remote energy sites
- Latency matters: Safety-critical equipment, commissioning checks, operator-attended machinery
- Data sovereignty is required: Defense, regulated industries, competitive environments
- Large sensor deployments need cost control: 20+ sensors where cellular data costs compound quickly
- Operational continuity is critical: Facilities that cannot tolerate monitoring gaps during internet outages
For a small deployment of 3–5 sensors in a well-connected facility, cloud-only processing may be perfectly adequate. For a marine vessel with 30 sensors and intermittent satellite connectivity, edge AI is not optional — it’s the only architecture that works reliably.
Conclusion
The shift toward edge AI in bearing condition monitoring is not a technology trend for its own sake — it’s a practical response to the real-world constraints of industrial IoT deployments. Connectivity is not guaranteed. Bandwidth is not free. Latency matters for some applications. Data sovereignty matters for some industries.
Edge AI moves the intelligence to where the data is generated, enabling real-time anomaly detection, bearing defect classification, and health scoring without cloud dependency. When paired with cloud synchronization for fleet analytics and model updates, it provides the best of both worlds: autonomous local monitoring with centralized fleet intelligence.
For a comparison of the wireless protocols that connect edge sensors to gateways, see our technical article on BLE vs LoRa vs LTE for bearing monitoring. Fault Ledger is one platform that implements on-gateway edge AI for bearing condition monitoring.