AI Computer Vision for Utilities CMMS: Drones, Gauges and Fleet-Scale Inspection

Utility maintenance operations produce a lot of imagery already: technician completion photos, drone flights over transmission corridors, inspection photographs from substation walkdowns, thermal images from transformer surveys, borescope footage from pressure vessels. Most of it gets attached to work orders and quietly ignored after the fact. Computer vision changes the economics of that imagery, and for utility operators specifically, the compounding value over a decade of records is where the strategic case sits. This post covers the utility-specific vision use cases that are production-ready today, the ones that are still fragile, and how utility maintenance leaders should think about deploying vision inside their operational technology and CMMS environments.

Scope note: this article is the utilities-CMMS variant of the broader Computer Vision for FM and CMMS hub. The commercial FM variant covering tenant complaint triage, statutory PPM evidence and building envelope inspection is at the FM computer vision pillar. For the SCADA-adjacent NLP variant, see the utilities CMMS NLP pillar.

Transmission-line and pipeline drone inspection

The single most operationally consequential vision use case in utility maintenance today is drone-based inspection of linear infrastructure. Transmission lines, distribution networks, pipeline corridors, canal networks, railway assets all share the same operational reality: long, physically remote, human inspection is expensive and slow, condition changes matter, and a small drone with a good camera plus a competent vision model can inspect kilometres per hour.

For transmission-line operators specifically, vision-assisted drone flights detect insulator damage, conductor wear, corona discharge patterns (with UV cameras), vegetation encroachment on the right of way, structural damage to towers, corrosion on steel work, and animal or bird interference. What used to be a linesman climbing a tower to visually inspect is now a drone circling with vision inference happening either on the drone or streamed to a ground station. The linesman only climbs when the drone flags something worth climbing for.

The compounding value over years is where it gets interesting. Fleet-wide condition data, trended quarterly against historical baselines, exposes deterioration patterns that single-flight inspections cannot see. Reliability engineering on utility linear assets moves from anecdotal to evidenced in ways that would have been impossible a decade ago.

Pipeline corrosion and structural detection

Water utilities, oil and gas operators, and district-energy operators all run large networks of pipe infrastructure that periodically requires visual inspection. Traditional methods (visual walkdowns, in-line inspection tools, cathodic protection surveys) produce data at intervals that leave large blind spots between inspections. Vision-assisted drone or crawler robot inspection fills those gaps.

Segmentation models detect corrosion severity, coating failures, mechanical damage, third-party interference (excavations, vehicle impact), and vegetation intrusion on above-ground pipe sections. Buried-pipe inspection is more constrained, but for the surface-mounted or exposed portions of utility networks, the vision layer is genuinely productive today.

On projects I have delivered to utility clients, the operational pattern that works consistently is drone-first, then trend, then dispatch. Drones fly the network on a schedule. Vision models produce a ranked defect list. Trend analysis compares against prior flights. Only defects that trend worsening or reach a threshold generate a work order in the CMMS. The reliability engineer's time compounds because they are looking at the ranked exceptions, not the whole network.

Pressure vessel and rotating equipment crack detection

For refining, chemical, petrochemical and district-energy operators, pressure vessels and critical rotating equipment carry serious consequence-of-failure implications. Statutory inspection regimes are demanding, and even between inspections, condition can change in ways that need catching early.

Vision-assisted inspection of pressure vessels covers external surfaces, seam welds, and (with borescope integration) internal surfaces. Segmentation models trace crack outlines, flag orientation and branching patterns, and score severity in ways that align with engineering assessment practice. The output is not a substitute for the qualified engineer's judgment, but it accelerates the engineer's work by pre-ranking findings across a fleet of vessels.

The same pattern applies to rotating equipment casings, structural steelwork on cooling towers, boiler externals, and any other high-consequence asset where visual condition matters and the fleet is large enough that human-only inspection is a bottleneck.

Legacy gauge digitisation

A common utility operational reality: a large installed base of analogue gauges (pressure, temperature, flow, level, running-hours counters) that are not connected to SCADA. Operators walk rounds with clipboards or tablets, recording readings manually. This is a labour cost, an error source, and a barrier to trending analysis.

Vision-based gauge reading digitises this workflow at a fraction of the cost of retrofitting each gauge with a digital transmitter. The operator points the phone at each gauge; the vision model records the reading; the data lands in the historian or CMMS; the operator confirms exceptions. For utility operations with hundreds of legacy gauges, the payback is measured in months, and the improvement in data quality (from opportunistic hand-transcribed readings to consistent digital records) supports downstream reliability analysis that was previously impossible.

Thermal imaging on transformers and switchgear

Thermal imaging is one of the most established pre-vision AI condition monitoring practices in utility operations. Bushings, transformer tanks, switchgear connections, motor bearings all reveal condition through their thermal signature. What has changed with modern vision is the analysis layer.

Traditional thermal-imagery analysis requires a qualified thermographer to interpret each image against a temperature reference and asset context. Modern vision layers automate the first-pass analysis: identify the asset type, locate the thermal anomaly, compare against baseline temperature signatures for that asset class, flag anomalies for the thermographer to confirm. The thermographer's role shifts from inspecting every image to reviewing the flagged exceptions.

Combined with drone-mounted thermal cameras and periodic overflights, this makes fleet-wide thermal-condition trending operationally practical. A substation or plant of hundreds of critical assets can be thermally inspected quarterly at costs that make the practice broadly adoptable rather than reserved for the most critical assets.

SCADA and vision cross-referencing

The most sophisticated utility deployments cross-reference computer vision findings with SCADA data streams. A vision-detected temperature anomaly on a transformer bushing is cross-checked against the SCADA-measured loading pattern; combined, the anomaly is either the routine consequence of high load or a genuine deterioration signal.

This cross-modal integration is where the reliability payoff compounds. Vision alone produces candidates. SCADA alone produces load data. Together, they produce contextualised findings that either survive investigation or filter out as expected. The maintenance planner's work becomes progressively higher-signal as this integration matures. (For the underlying framework, see the AI copilot for utilities CMMS pillar.)

Where vision is still fragile in utility contexts

• Poor-visibility conditions: fog, rain, night operations, dust-laden air (arid environments), snow accumulation. Vision degrades in these; plan around it either with weather-window scheduling or with sensor-fusion approaches (thermal, LiDAR, radar).
• Buried or enclosed assets: buried pipe, cable in ducts, sealed switchgear, enclosed bearings. Vision helps with what is visible externally; complementary techniques (in-line inspection tools, ultrasound, vibration analysis, partial discharge monitoring) remain essential.
• Regulatory and safety integration: drone flights over transmission corridors need airspace approval, competent pilots, safety protocols. Vision integration into safety-critical inspection regimes needs qualified sign-off; the vision layer accelerates but does not replace the qualified inspector.
• OT-IT integration constraints: pushing vision-derived findings into a SCADA-controlled operational environment needs the same cybersecurity discipline as any other cross-boundary integration. Do not shortcut the OT security review.

The business case for utility operators

Five benefits utility operators typically realise from a focused computer vision programme:

• Inspection productivity at fleet scale: drone-based vision-assisted inspection covers linear infrastructure at a fraction of the cost of foot inspection, with more consistent coverage.
• Trend analysis becomes real: fleet-wide condition data, indexed and trended over years, exposes patterns that human inspection cannot see across the fleet.
• Legacy instrument digitisation: reading analogue gauges via vision cheaply digitises data streams that would otherwise require sensor retrofit spend.
• Faster regulatory response: vision-derived condition records support regulatory submissions and inspections with structured evidence.
• Safer inspection regimes: hazardous zone inspections shift from human-in-danger to robot-or-drone-first, keeping people out of the highest-risk work.

Where utility maintenance leaders should start

1. Pilot gauge digitisation first. Concrete, measurable, low-risk, immediately valuable for reliability trending. Payback in months, trust built quickly with the team.
2. Then extend to drone-based inspection of one linear-infrastructure asset class (transmission corridor, one pipeline network, one substation fleet). Prove the pattern at defined scope before expanding.
3. Engage OT security and airspace-regulator conversations early. These are not procurement decisions; they are engineering and compliance decisions with lead times.
4. Defer safety-critical generative use cases. Vision-derived findings as candidates: yes. AI-generated safety procedures or isolation instructions: no. Draw the line clearly and communicate it to the team.

For the broader utility AI frame, see the AI copilot for utilities CMMS pillar, the utilities NLP pillar, and the underlying asset criticality classification pillar.

Final thoughts

Computer vision in utility maintenance has moved decisively from research to production. The vendors are shipping. The drones fly. The models work well enough on the important use cases (transmission-line inspection, gauge reading, pipeline surface corrosion, thermal-anomaly triage). Where utility operators go wrong is treating vision as a technology purchase rather than an operational-and-cultural shift.

The operations that get value from vision are the ones that scope carefully, integrate with SCADA and CMMS deliberately, respect the OT security boundaries, and design the human-in-the-loop pattern so that the qualified engineer's time compounds rather than gets replaced. The technology is not the challenge in 2026; the design and change management are. Get those right and vision earns its keep for a decade of operational value.

Related reading: Computer Vision hub, AI copilot for utilities, NLP for utilities, asset criticality classification, AI governance for enterprise operators.