Buried Assets, Blind AI: Why Predictive Maintenance Works Differently for Linear Infrastructure

When vendors claim their AI can predict asset failures six to twelve months in advance with over 85% accuracy, they are almost certainly talking about rotating equipment — pumps, compressors, turbines, motors — where continuous vibration and thermal sensors give the model a direct window into what the asset is doing right now.

More than 70% of a typical water utility's assets are buried underground. They have no sensors. They are invisible to the real-time data pipelines that power the new wave of predictive maintenance platforms. The 2026 WaterWorld industry benchmark found that condition data for high-risk assets — asbestos cement mains, cast-iron distribution sections, ageing sewer rising mains — is sparse or outdated at most utilities. Only three in ten are satisfied with their asset management program. AI adoption, the same benchmark found, remains almost nonexistent.

This is not a technology gap. It is a reasoning gap. And it affects every operator who manages linear infrastructure alongside the facilities that serve it.

Linear assets live in networks, not hierarchies

Discrete assets — pumping stations, substations, compressor stations, manufacturing lines — sit in hierarchies. A facility contains systems; systems contain components. Sensor data flows up the hierarchy. When a bearing on pump P-101 starts running hot, a vibration sensor catches it, the AI flags the anomaly, and a work order is generated. The logic is tractable: one asset, one sensor stream, one failure mode to classify.

Linear assets don't work this way. A buried water main, a gas distribution network, a rail corridor, a transmission line — these are topological objects. Failure modes are not local; they propagate. A degraded pipe section doesn't just affect that section — it creates pressure stress on adjacent connections, risks cascade failure at the nearest junction, and determines which communities lose supply first if it ruptures. The question isn't "when will this asset fail?" It's "what does a failure here propagate to, and in what order?"

To answer that question, you don't need a sensor on the pipe. You need a model of the network. You need GIS-native topology, hydraulic or load-flow modelling, historical failure records cross-referenced against soil composition, age, material type, and operating pressure. You need to reason from indirect evidence — pressure variance at monitoring points, flow anomalies, acoustic correlations — and trace likely fault locations through the graph.

This is what linear asset management actually requires. It is a fundamentally different computational problem from predictive maintenance on instrumented equipment, and conflating the two leads organisations to buy tools optimised for one that simply do not perform on the other.

When the same operator has both

Most utilities, transit authorities, and energy companies have both asset worlds inside the same organisation — and this is where a quiet compounding failure takes hold.

The pumping stations serving your pipeline network are discrete assets. Your water treatment plant is a discrete asset. Your maintenance depots, substations, and control facilities are discrete assets. They generate telemetry. They benefit from AI-powered failure prediction that works well on instrumented equipment.

But they don't exist in isolation from the linear network they serve. A cascade failure in the distribution network — the invisible, buried part — directly loads those downstream facilities. If section 7B of your main degrades and flow rates drop, the pumping station at the end of that run is operating outside its normal parameters. The AI watching the pump doesn't know that. It isn't looking at the pipe.

Most operators run separate systems for these two worlds: a GIS-adjacent platform for the network, a CMMS or traditional EAM for the facilities. The failure modes of one don't appear in the risk model of the other. When an incident happens — a burst main causes downstream pump failure, a degraded rail section creates abnormal loads on depot equipment — the post-incident review must stitch together records from two systems that were never designed to share a data model.

The compliance clock is ticking

New regulations are arriving on top of this structural problem. ISO 18379-1:2026, published in May, requires rail operators to document safety and integrity considerations at every stage of infrastructure lifecycle — from initial design through decommissioning. NERC CIP updates effective in 2026 extend cybersecurity and asset inventory requirements deeper into grid infrastructure than previous versions covered. US utilities are simultaneously facing a $208 billion investment shortfall in ageing transmission assets built in the 1950s and 1960s — assets for which comprehensive condition records may never have existed.

Demonstrating compliance when your asset data lives in two disconnected systems — with inconsistent taxonomy, separate audit trails, and no shared risk model — is a quarterly reconciliation exercise that produces documentation, not assurance. Regulators increasingly ask not just "what is the condition of your assets?" but "how did you determine that, and what is your evidence?" A decision log tracing every risk prioritisation back to its source data is becoming the expected answer. A disconnected EAM stack cannot produce one.

The right design for both

UniEAM's approach to linear assets is built around network-aware reasoning: GIS-native topology models, cascade failure graphs, and probabilistic risk scoring that draws on indirect evidence rather than requiring sensor coverage on every buried segment. The system infers probable fault locations from operational data — pressure differentials, historical break rates by material and age cohort, proximity to high-consequence zones — and produces risk prioritisations with a full reasoning trace.

The discrete asset module runs the complementary logic — hierarchy-aware reasoning across ISO 14224 failure taxonomy, OEE-grade telemetry integration, and parent-child failure mode propagation through facility hierarchies. Both modules share the same data model and the same agent layer. When the pumping station at the end of your distribution main shows anomalous behaviour, the system can query whether upstream pressure variance in the linear network is a plausible contributing cause — and log that inference as a traceable reasoning step.

That is what it means to manage both asset types as one portfolio, rather than two adjacent problems with a manual export in between.

The buried 70% of your infrastructure isn't going to get sensors any time soon. But it doesn't need to — if the reasoning approach is right.

Curious how UniEAM handles both? Request a demo.