Predictive Maintenance Across a Distributed Fleet

When the asset sits at the other end of the country

Pumping stations and compressors often sit where no one passes by every day. Spread across dozens or hundreds of sites, each unit nearly identical, each one unremarkable on its own. When one fails, it gets expensive, with unplanned downtime, emergency call-outs, and knock-on damage to downstream components. The classic reflex is to ship all sensor data into a central cloud and analyze it there. In practice that runs into bandwidth limits, shaky cellular connectivity, latency, and the question of who actually owns the operating data.

Predictive maintenance on a single asset with classic TinyML is a solved problem. The open question is scaling across a distributed fleet. How does each unit detect its own fault locally, and how does the whole fleet grow smarter together, without raw data ever leaving the site? This is exactly where an Edge Agent comes in, an AI agent on the machine that decides on site, plus a fleet layer that shares what it has learned.

Detection belongs on the machine, not in the cloud

The first line of defense runs entirely as on-device AI, with no constant connection. Every asset runs the same engine as a single binary, whether on a small Linux box, an embedded controller, or x86 hardware in the cabinet. Redeploy, not rewrite.

Detection itself is deliberately sober. Threshold triggers with deadband catch the obvious cases, such as pressure, temperature, vibration, or current draw outside the expected band. On top sits classical machine learning for anomalies that no fixed limit describes, such as slow drift or unusual patterns in the load profile. Both run deterministically inside the workflow graph, which is the actual program. The AI is just one node in it, inspectable, bounded, traceable. No black-box model acting on its own.

This matters in industry. The agent never sits inside the real-time control loop. The deterministic control of the pump keeps running undisturbed. The Edge Agent observes, assesses, and at most writes buffered parameters back, following the “agent as a tool” pattern, never as the controller.

The AI agent reports and explains

A bare alarm reading “anomaly at station 47” does little for the on-call crew. The Edge Agent goes a step further. When a threshold or the ML model fires, a local language model (a small SLM with 1 to 3 billion parameters, served by a local inference server right next to the engine) composes a concise explanation of which signal is off, which direction it is drifting, and which plausible cause it points to.

This assessment is produced on site, with no cloud call. A structured message then goes out over MQTT to the control level, including plain text, not just a raw value. The service technician knows whether to head out immediately or whether the next routine visit will do. That is precisely what separates local AI in industry from a bare sensor stream, namely context instead of a flood of data.

Federated learning lets the fleet learn together

This is where distribution pays off. Over its lifetime, any single station sees only a handful of genuine faults. A rare bearing problem on one pump is barely statistically meaningful locally, but across two hundred identical units it very much is.

Federated learning solves this without centralizing raw data. Each asset keeps training on its own operating data, which never leaves the site. Only the model updates (the learned weights, not the measurement series) flow back, get aggregated across the fleet, and are pushed out again as an improved model. A fault signature that appears for the first time at one location then sharpens detection on all the others.

It is also GDPR-friendly. Sensitive operating and consumption data stays physically on site, and only what is needed to learn travels to the center. An Agent Registry and fleet orchestration keep track of which unit is running which model version.

A scenario with two hundred nearly identical pumping stations

An operator runs around two hundred pumping stations across a wide service area. The stations are technically nearly identical, yet sit in very different places, some with stable connectivity, many reachable only over variable cellular links.

Here is how an Edge Agent could solve this case. At one station, a pump’s current draw slowly begins to rise while delivery output stays the same. No single value breaches a hard limit, yet the local ML model still flags the drift as atypical. The agent explains on site that the pattern fits the onset of mechanical friction and issues an early warning, days before a classic threshold would have tripped. The signature flows into the fleet as a model update. Weeks later a second station shows the same pattern, and the system recognizes it instantly.

How it works at the edge

The flow on each unit, in concrete technical terms.

• Acquisition and triggers. Sensor values come in over analog inputs and serial interfaces as first-class nodes. Threshold triggers with deadband and a classical ML model continuously check for anomalies, all offline-capable inside the deterministic graph.
• Local explanation. On a hit, an on-device SLM produces a concise cause assessment via the local inference server. Durable on-device memory holds history and context.
• Reporting. A structured message goes over MQTT to the control level, and the asset’s control loop keeps running untouched.
• Federated update. The locally retrained model delta is aggregated across the fleet (without raw data) and redistributed as an improved version.

The security architecture stays sober throughout, with local execution, container isolation, and the auditable, bounded graph. Signed image distribution and on-device execution are designed from the ground up for requirements such as the EU Cyber Resilience Act, with CRA readiness as an architecture property, not a label.

Key Takeaways

• Predictive maintenance across a distributed fleet is not a cloud scaling problem but a question of local detection plus shared learning.
• Each Edge Agent detects anomalies on site (threshold plus classical ML), explains them with an on-device SLM, and reports over MQTT, without touching the real-time control loop.
• Federated learning makes the fleet better together, with models improving across devices while raw data stays on site (GDPR-friendly).
• The same engine runs as a single binary on whatever hardware sits in the station, redeploy, not rewrite.
• From Hardware to Intelligence. Sober detection at the machine, shared knowledge across the whole fleet.