Reactive maintenance is expensive. A piece of critical manufacturing equipment fails, the production line stops, engineers get called in, parts are sourced under time pressure, and the actual repair is often the cheapest part of the incident. The lost production, the emergency labour rates, the expedited parts shipping: those are the numbers that make plant managers wince. Planned maintenance is better — scheduled shutdowns, replaced components on a calendar — but it’s also wasteful, because you’re replacing parts that might have months of useful life left.
Predictive maintenance sits between these two. You monitor equipment condition continuously, detect when something is drifting toward failure, and intervene before the failure occurs. You replace the bearing when the vibration signature tells you it’s deteriorating, not when the calendar says it’s time or when it breaks. The concept has been around for decades. What’s changed is the availability of edge AI hardware and software that makes it economically viable at the individual machine level.
Why Edge, Not Cloud
The obvious question is: why not just stream sensor data to the cloud and run the analysis there? Three reasons dominate in industrial environments.
Connectivity is not guaranteed. Many manufacturing facilities have factory floors with patchy WiFi coverage, machinery that generates electrical interference, or simply areas where reliable uplink is impractical. An analysis system that depends on constant cloud connectivity is too fragile for production use.
Latency matters for intervention. If a vibration anomaly indicates a bearing is about to fail catastrophically, you want an alert in seconds, not minutes. Round-trip cloud processing adds latency that is unacceptable for the highest-stakes scenarios. Edge inference completes in milliseconds.
Data volume is prohibitive at cloud scale. A single vibration sensor sampling at 10-20 kHz generates enormous data volumes. Streaming raw sensor data from dozens of machines to the cloud is expensive and often unnecessary — you want the insight (anomaly detected on motor 7) rather than every raw sample.
Edge processing solves all three: the analysis runs locally, latency is minimal, and only alert events or aggregated summaries travel upstream.
What Sensors Measure
Predictive maintenance draws on several sensor modalities, often in combination.
Vibration is the workhorse. Accelerometers attached to rotating equipment — motors, pumps, compressors, fans — measure the characteristic vibration signatures of the machine during normal operation. As bearings wear, as imbalances develop, as alignment degrades, the vibration signature changes in characteristic ways. These changes are often detectable weeks or months before the machine fails. Vibration-based condition monitoring has been used in high-value industrial environments for years; edge AI makes it accessible at smaller scale.
Temperature is simpler to measure and tells you something useful: equipment running hotter than normal is usually equipment that is under abnormal stress. Motor windings warming above design temperature indicate either overloading or inadequate cooling. Thermal imaging (now available in compact, affordable infrared modules) can map temperature distribution across components and detect hotspots that point sensors would miss.
Current draw on electric motors contains information about mechanical load and efficiency. A motor working harder than normal to drive the same load is a motor where something is wrong — increased friction, misalignment, a developing fault in the driven equipment. Current-based monitoring requires no physical attachment to rotating parts, which is an operational advantage in environments where sensor installation is difficult.
Acoustic emission picks up ultrasonic frequencies associated with friction, stress fractures, and leakage. These are inaudible to human ears but detectable with appropriate sensors, and they can identify problems that vibration analysis misses at early stages.
The AI Model Side
The analysis that runs at the edge is typically anomaly detection rather than specific fault classification. The model learns what normal looks like for a specific machine in its operating context — not “what does normal vibration look like for a motor of this type” but “what does normal vibration look like for this specific motor in this factory, driving this load, at these ambient temperatures.” That specificity matters because every installation is slightly different.
During a commissioning period, the model collects baseline data and establishes normal operating profiles. After that, it looks for deviations: vibration frequencies that weren’t present before, temperature patterns that differ from historical norms, correlations between sensor readings that have changed. When deviation exceeds a threshold, an alert is generated.
The models deployed at the edge for this purpose are relatively lightweight — they’re not large language models or computer vision transformers. Time-series anomaly detection can be done effectively with compact neural networks or even classical statistical methods like CUSUM or ARIMA with machine-specific parameter fitting. This means they run well on embedded hardware: ARM Cortex-based edge compute devices, Raspberry Pi Compute Module 4 in an industrial enclosure, or purpose-built industrial edge gateways from vendors like Moxa, Advantech, or NEXCOM.
Larger edge AI accelerators (Hailo, Intel Neural Compute Stick, NVIDIA Jetson) are used when the application involves higher-dimensional data or multiple simultaneous sensor streams that benefit from GPU acceleration. For most vibration and temperature monitoring applications, a capable CPU is sufficient.
Implementation Patterns
The typical deployment has three layers. Sensors transmit data (over wired RS-485, 4-20mA loops, or wireless protocols like Bluetooth LE or WirelessHART) to a local edge gateway or industrial PC. The edge device runs the anomaly detection model, maintains a rolling history, and generates alerts. Alerts and aggregated statistics are pushed to a central maintenance management system — cloud-based or on-premises — where maintenance engineers act on them.
Connectivity between edge and central systems can be intermittent without losing functionality, because the anomaly detection logic runs entirely locally. The edge gateway can buffer alerts during connectivity outages and sync when the connection is restored.
Alert quality — specifically, avoiding alert fatigue from too many false positives — is the main operational challenge. A system that cries wolf constantly gets ignored. Model tuning after deployment, based on feedback from maintenance engineers about which alerts led to real findings and which didn’t, is important and often underestimated in project planning.
The Business Case
The ROI calculation for predictive maintenance is usually compelling for high-value equipment or processes where downtime is expensive. A single avoided failure on a critical production line component can pay for the monitoring infrastructure several times over. The difficulty is quantifying this before the first avoidance event, which makes upfront justification harder than it should be.
A practical approach is to start with the equipment where failure is most expensive and most predictable by condition monitoring — high-speed rotating machinery, pumps in critical processes, compressors — and treat the first year as a data-gathering phase. The baseline data you collect has value beyond monitoring: it tells you what your equipment is actually doing, which often reveals inefficiencies that are worth addressing independently of the failure detection goal.