AI Computer Vision for FM and CMMS: Image-Based Inspection and Asset Recognition

Every maintenance work order attached with a photo is a data point that a computer vision model could interpret. Every quarterly building inspection captured in photographs is a data set. Every drone flight over transmission lines produces imagery that used to require an engineer's hours to review and now can be pre-triaged in minutes. Computer vision is not a niche AI capability inside maintenance operations any more. It is a foundational layer that reshapes the work of inspecting, categorising, and acting on physical assets. This post covers what computer vision actually does in FM and CMMS contexts, where it works reliably today, where it is still fragile, and how buyers should think about it as they plan the next 24 months of their operational technology roadmap.

What computer vision does in maintenance operations

Computer vision, in the modern AI sense, is the family of models that interpret visual inputs (photos, video, thermal images, LiDAR, hyperspectral scans) and return structured information: what the object is, where it is in the frame, whether it exhibits an anomaly, and how confident the model is in its judgment. The underlying technology has matured from convolutional neural networks (CNNs) through to vision transformers and multimodal foundation models that can be prompted with natural language.

For FM and CMMS, three model patterns matter most. Object detection identifies and locates specific items in an image (an AHU unit, a pressure gauge, a hard hat, a leak). Segmentation outlines the exact boundaries of an anomaly (the shape of a crack, the extent of corrosion). Classification assigns categories to the whole image or a region (this asset is in poor condition, this technician is not wearing PPE). Modern platforms combine all three, often chained with language-model reasoning to produce a human-readable inspection summary.

Where the models run is a design decision. Cloud-hosted inference gives access to the largest models and the newest capabilities but requires connectivity. Edge inference (on the phone, camera, or a small on-site appliance) works offline and preserves privacy but is bounded by device compute. Most production deployments use a hybrid pattern: edge for fast pre-filtering, cloud for anything the edge is uncertain about.

AI-powered inspection using photos

The workhorse use case. A technician on a mobile app photographs an asset during routine work. The vision model interprets the image and returns structured findings: the asset type is confirmed against the CMMS record, any visible anomalies (rust, dents, damage, leaks) are flagged with severity, and the model drafts a short narrative summary the technician can accept or edit. The completed inspection lands in the CMMS with structured data attached, not just a photo the future analyst has to interpret.

The change in workflow is subtle but consequential. Traditional photo-attached inspections rely on humans to look at the picture later. Computer-vision-enabled inspections extract the signal at the point of capture, making the inspection itself the data source. Fleet-scale reliability analysis becomes possible in ways that photo archives never enabled, because every image is now structured.

Leak, corrosion and crack detection

Three specific anomaly types where computer vision has moved from research to production:

• Leak detection: dripping fluids, wet patches, staining, discolouration around fittings. Models trained on this pattern can flag potential leaks with useful accuracy in typical mechanical rooms, plant rooms and pipe galleries. Water leaks are easier than gas or steam. Thermal-imagery pairing lifts accuracy significantly.
• Corrosion detection: surface rust, pitting, blistering paint, oxide bloom. Particularly valuable for outdoor assets (tanks, pipe networks, structural steel, HVAC roof units) where visual inspection is standard practice. Models can rate severity from image alone with reasonable accuracy, and the change over time (comparing photos across quarters) is where the compounding value sits.
• Crack detection: cracks in concrete, steel welds, pressure vessel surfaces, façades. Segmentation models trace the crack outline and can flag orientation, length and branching, which are the metrics engineers actually use to judge severity. Widely deployed in bridge inspection, dam monitoring, and industrial pressure-vessel walkdowns.

Each of these can be delivered by a smartphone-based mobile app the technician carries, by a fixed camera monitoring a critical asset, or by a drone or robot flying the inspection route. The underlying model is the same; the deployment surface is what varies.

Equipment identification and nameplate OCR

The other half of inspection work is knowing which asset you are looking at. Computer vision models trained on equipment imagery can identify pumps, motors, valves, switchgear, panels, chillers, AHUs and dozens of other equipment classes from a photograph, often with the specific manufacturer and rough model. This closes a common data-quality gap in maintenance operations, where technicians know what they are looking at but the CMMS asset tag was worn or missing.

The complementary capability is nameplate OCR. Every industrial asset carries a nameplate with make, model, serial number, rating and other metadata. Reading these plates has traditionally been manual, error-prone, and one of the most tedious parts of asset onboarding. A vision model photographing the plate can extract the structured data reliably. For asset onboarding projects covering thousands of items, this alone can shave months from a data-migration timeline. (Broader data-migration context in the CAFM data migration pillar.)

Reading gauges automatically

A large installed base of industrial gauges (pressure, temperature, flow, level, hours-run counters) is not connected to any SCADA or IoT layer. They are read by operators walking rounds with clipboards. Computer vision has become good enough to read these gauges reliably from a photograph. A round-book that used to be paper is now a mobile app; the operator points the phone at each gauge, the vision model records the reading, the data lands in the CMMS or historian.

The economics matter. Retrofitting a fleet of analogue gauges with digital transmitters is expensive and disruptive. A vision-based digitisation layer costs a fraction and lifts the data quality of the existing infrastructure without touching the physical asset. For asset-intensive operations with hundreds of legacy gauges, this is one of the fastest-payback vision use cases available.

PPE detection and site safety inspections

Safety compliance monitoring is another use case where computer vision has crossed the production threshold. Fixed cameras at site entrances and inside plant areas can detect whether personnel are wearing the required PPE (hard hats, high-visibility vests, safety glasses, respirators depending on the zone). Real-time alerts on non-compliance shift the safety function from post-incident investigation to preventive intervention.

For FM operations, the same technology handles broader site safety inspections: unauthorised access, blocked emergency exits, spills, hazardous storage violations, tripping hazards. Combined with mobile-based inspection workflows, this turns the periodic safety walkdown from a checklist exercise into a continuous data stream that surfaces issues in near-real-time.

The honest caveat: PPE detection specifically has faced criticism when deployed without deliberate policy design. False positives create friction with workers. Focus on aggregate compliance trends and anomaly surfacing rather than individual-worker enforcement. This is a case where the technology can outrun the operational policy, and the policy has to catch up before deployment lands well. See AI governance for enterprise operators for the broader framing.

The future direction: autonomous inspection

Three deployment surfaces will reshape how inspection actually happens over the next 24 months.

Drones and unmanned aerial vehicles: fixed-route drone inspections of transmission lines, pipeline corridors, solar farms, wind turbines, building façades and roof estates. Software-defined flight paths, autonomous take-off and landing, and vision-based anomaly detection pipe findings straight into the CMMS. What used to take a two-person team a full day now happens in an hour, with more complete coverage.

Ground-based mobile robots: quadruped robots (Spot, ANYmal), tracked robots, autonomous ground vehicles. Deployed in substations, plant rooms, tunnel networks, and hazardous zones where sending a human is either dangerous or inefficient. Vision-model inference happens on-robot or streams to a cloud endpoint. Adoption is growing in utilities and heavy industry.

Fixed cameras with intelligent analytics: strategically placed cameras monitoring critical assets 24/7, with anomaly-detection models flagging changes. Costs less than robots and drones, works well for stationary high-consequence assets (transformers, pressure vessels, critical rotating equipment). Increasingly bundled with SCADA and CMMS platforms as vendors integrate the vision layer natively.

The pattern across all three is the same: the human role shifts from performing the inspection to reviewing the exceptions the vision layer surfaces. The technician's time compounds because they only look at what the model has flagged, which is where their expertise adds value.

Honest limitations to plan around

Computer vision in FM and CMMS is genuinely useful today, but it is not magic. Four limitations worth being clear about at the design phase:

• Data-labelling costs: models need training data specific to your asset base, environment, and lighting conditions. Generic pretrained models get you started; production accuracy needs domain-specific labelling. Budget for this as a first-year investment, not as an afterthought.
• False positives are operational friction: a model that flags every possible anomaly overwhelms the human reviewer. A model that misses genuine anomalies undermines trust. Tuning the precision-recall trade-off is not a one-time exercise; it is an ongoing operational discipline that matures over 12 to 24 months of production use.
• Camera and image quality matter: poor lighting, motion blur, wet lenses, oblique angles all degrade model performance. The physical acquisition layer (cameras, lens types, lighting rigs, gimbals for drones and robots) is often the actual bottleneck, not the AI model. Invest here alongside the AI investment.
• Physical constraints of visual inspection: vision cannot see inside a pump bearing, through a sealed enclosure, or beneath insulation. Where the failure mode is invisible externally, no amount of visual AI helps. Pair visual inspection with vibration analysis, ultrasound, thermal imaging, or oil sampling for a complete condition-monitoring picture.

The business case, honestly

Five benefits organisations typically realise from a focused computer vision rollout in maintenance operations:

• Inspection productivity: automated pre-filtering compresses inspection time from days to hours for large asset fleets, and lets the human expert focus only on exceptions.
• Data quality at source: photos that used to be untagged now carry structured metadata (asset ID, anomaly type, severity), making downstream reliability analysis credible for the first time.
• Faster asset onboarding: nameplate OCR alone can cut new-asset registration time by 60 to 80 percent for large data-migration exercises.
• Safer operations: PPE and site safety monitoring shifts the safety function from lagging (incident investigation) to leading (preventive intervention).
• Legacy-instrument digitisation: reading analogue gauges via vision cheaply digitises data streams that would otherwise require expensive sensor retrofits.

The honest costs to plan for: initial data-labelling investment (weeks to months of specialist effort), camera and acquisition-hardware capex, ongoing model maintenance as the asset base evolves, and organisational change effort to move from "manual inspection with photos" to "vision-assisted inspection with human review of exceptions." None of these are trivial, but all pay back well when the operational use case is scoped correctly.

Where to start with computer vision

Three practical steps for a maintenance or facility leader thinking about this:

1. Pick a high-volume, low-risk inspection workflow first. Nameplate OCR for asset onboarding, or PPE compliance monitoring at a single site, or gauge-reading for a monthly round. Prove the pattern operationally before committing to broader rollout.
2. Solve the image-acquisition problem alongside the model problem. Cameras, lighting, mobile-app UX, and workflow integration matter as much as the AI. A model that never gets good images fails silently.
3. Design the human-in-the-loop pattern deliberately. Vision surfaces candidates; humans confirm and act. Getting this handoff clean is what separates production-grade deployments from expensive experiments.

For the broader AI-in-maintenance frame this sits inside, see the AI Copilot for FM and CMMS pillar and the NLP variants (FM NLP, utilities NLP). The FM-specific and utilities-specific spokes of this computer-vision pillar are covered in dedicated posts (linked below when live).

Final thoughts

Computer vision has crossed the threshold from research demo to production infrastructure in maintenance operations. Vendors are shipping capable products. Drone and robot platforms are becoming operationally viable. The barriers today are less about model capability and more about data labelling, image acquisition, and the organisational design needed to shift human roles from performing inspections to reviewing them.

Buyers who scope carefully (one high-volume workflow first, deliberate human-in-the-loop design, honest planning around labelling and camera investment) get real operational value within a year. Buyers who chase every vision use case at once tend to end up with expensive experiments and disillusioned teams. Sequence matters. Pick well, land the pattern, then expand.

Related reading: AI Copilot for FM and CMMS, NLP for facility management, RAG over CAFM data, AI governance for enterprise operators, CAFM data migration strategy.