Defect severity classification for weather-infrastructure assets works when each severity level is defined by the response it requires rather than by adjectives, when severity is judged on consequence and likelihood rather than on how bad a defect looks, and when severity is kept separate from the scheduling priority a work order carries. Four levels are usually enough, as long as each one states a response time, an escalation route and the evidence it needs. For radar domes, masts and remote enclosures, that shared framework then needs asset-specific acceptance criteria, a named sign-off authority, a criticality weighting so the review queue ranks by risk instead of by timestamp, and a translation layer for any legacy work-order field that only understands pass or fail.
The observing network is the physical end of AI engineering for weather and climate: radars, masts and enclosures whose condition decides whether the data reaching forecast systems can be trusted. The Met Office uses its station grading for record verification because the grading “tells us how much confidence we have in the data”. A vision model that flags corrosion on every mast in a network produces a list; the taxonomy and the ranking are what turn that list into work.
Defect severity levels defined by the response they trigger
A defensible defect severity taxonomy starts by separating the type of finding from its severity, because a single undifferentiated “defects” list makes minor observations look equivalent to conditions that need urgent intervention. Finding type says what kind of nonconformance the inspector saw; severity says how urgent it is and what could follow from it. Appearance is a poor guide: a large cosmetic defect can be harmless while a small defect in a critical location can signal a serious failure mode, so severity has to weigh what could happen if the condition remains and how likely that outcome is.
Words like “minor” and “serious” invite argument, while response requirements do not. FieldEagle’s guidance gives a worked definition of high severity: escalated to the responsible manager within four hours, assigned within one business day, and corrected or placed under an approved interim-control plan within seven days. A four-level scale built on that principle looks like this.
| Level | Defining response |
|---|---|
| Critical | Stop or isolate where required, apply immediate interim controls, notify designated management, obtain engineering review, raise an urgent corrective action |
| High | Escalate within a defined short period, assign an accountable owner, set a near-term due date, apply temporary controls and increase monitoring |
| Medium | A named owner and an enforced due date |
| Low | A routine corrective process |
Every severity level has to lead to a credible response. When only the top category gets action, inspectors start rating everything high, and severity inflation becomes predictable. Some conditions should skip the debate entirely: evidence of imminent structural collapse or exposed live conductors belong in pre-defined stop-work and emergency-escalation thresholds built into templates and workflows.
Categories for radar domes, masts and remote enclosures
The framework should be common across asset types while the technical thresholds stay specific to each asset; the same corrosion depth can be minor on a nonstructural enclosure and critical on a pressure-retaining component. The taxonomy should express the differences between radar domes, masts and enclosures in its acceptance criteria rather than in separate frameworks.
Radar domes. For radomes, damage has electrical as well as structural consequences. The FAA’s advisory circular on weather radar radomes states that any hole, regardless of size, can cause major damage because moisture enters the wall and causes internal delamination, with worse damage if it freezes. Enough moisture distorts the radiation pattern and seriously attenuates transmitted and returned signals. Repairs matter as well: even minor repairs can affect transmissivity, reflection and diffraction, and poor radome performance can look like faults in other parts of the radar system. Aviation radome repair guidance, as summarised by AviationPros, sets size thresholds above which a repaired radome should be electrically tested, such as core-penetrating repairs larger than four inches in their largest dimension, along with rules for multiple or overlapping repairs. Those figures come from aircraft radome guidance, so for a ground weather radar dome they are a template for the kind of threshold to define, with the dome manufacturer’s criteria supplying the numbers. The practical rule for the taxonomy: a breach in a radome skin should never default to low severity, and a repair history belongs in the record.
Masts and towers. PG&E’s guide for assigning priority codes grades steel-structure conditions such as moderate damage to main structural members, a missing bolt on a single-bolt connection to a critical member, cracking, pack rust at joints and rust or wear expressed as percentage material loss, and sends out-of-plumb structures to engineering for evaluation. The same guide distinguishes a single missing bolt in a multi-bolt connection from a missing bolt on a single-bolt critical connection, which is exactly the finding-type versus severity split in practice.
Remote enclosures. For an enclosure, severity usually follows what the enclosure protects: corrosion on a cabinet skin matters little until it threatens sealing, mounting or the instruments inside. The Met Office station scheme is a useful reference for what sits around the enclosure, since it assesses instrument conformance, installation, exposure and observer competence.
Visual inspection alone has known limits for judging structural severity. The IEA Wind Task 46 erosion work notes that depth of damage is particularly hard to categorise from inspection imagery, and the FHWA’s study of bridge inspection concluded that in-depth inspections using visual inspection alone are unlikely to identify the specific defects they are prescribed for. A taxonomy should therefore name which categories require a follow-up measurement or test before the severity is confirmed.
If an organisation already runs a visual-inspection taxonomy for energy or transmission assets, extending it with weather-infrastructure categories is the faster and more consistent route. Borrow the framework, not the thresholds: generic severity scales borrowed from other organisations are the third most common cause of failed criticality programmes. Where no base taxonomy exists, the technical criteria should be commissioned from a qualified structural or inspection authority and approved before anyone maps them to ticket priorities.
Who should sign off a defect severity taxonomy?
Sign-off belongs to a named, qualified authority, and the approval has to be recorded so it can be produced when a classification is challenged. Established schemes show what that looks like. Under DMRB CS 450, inspection records for highway structures are reviewed and authorised by the Authorising Engineer within six weeks of upload. PG&E’s priority-code guide was developed, and is regularly updated, by a cross-functional team of field personnel and subject matter experts. In SCE’s transmission programme, Priority 2 discrepancies are reviewed and approved by a Gatekeeper in the work management system. Criticality scoring follows the same logic: one function scoring alone produces results that cannot be defended in audits, and category weightings should be documented, approved by leadership and reviewed annually.
The signature is best split in two: technical acceptance criteria signed by the qualified engineer responsible for the asset class, and the response definitions and priority mapping signed by the operational owner who answers for maintenance response. Revisions need the same discipline. When a reviewer upgrades or downgrades a finding, the record should keep the original and revised classification, who changed it, when, why, the supporting evidence and the approval. When the taxonomy itself changes, the organisation has to decide whether historical findings keep the old value or are mapped into the new structure, and often both are needed: the original for auditability, a standardised value for trend reporting. Keeping a taxonomy and the models trained against it aligned through those revisions is its own problem, covered in keeping defect taxonomies and inspection pipelines robust through change.
Defect severity and priority answer different questions
Severity and priority are two independent axes, and every triaged defect needs both. Severity measures how badly the defect affects the system when it acts; priority is a scheduling decision, influenced by severity but also by business context such as regulatory deadlines. In maintenance terms, priority decides how the work competes for labour, materials and shutdown access. A high-severity finding may need immediate temporary control while its permanent repair waits for a planned shutdown, and a medium-severity finding on an easily reached asset may be fixed first; the severity stays high either way.
Temporary controls do not erase the defect. The record should show the original severity, the interim control, the residual risk and the permanent action separately, because downgrading the original severity makes later reviewers think the condition was never serious.
This is where the utility schemes can look contradictory. SCE allows a lower priority rating once a temporary repair stabilises a Priority 1 condition, and California’s GO 95 Rule 18 likewise lets a utility repair temporarily and reclassify to a lower priority level. Read with the severity/priority split in mind, there is no conflict: those rules move the scheduling priority, while keeping the field severity in the inspection record is the separate practice FieldEagle recommends. EPRI’s blade guidance treats a repaired Category 5 defect as Category 3 until operating experience shows the repair holds.
Mapping corrosion, cracking and loose fittings onto ticket priority levels
Corrosion, cracking and loose-fitting findings map to ticket priorities in two steps: the asset-specific criteria place each finding in a severity level, and a signed-off table maps each severity level to the priority codes and response windows the ticketing system already uses. Service levels belong to the priority, not to the severity: response targets are configured in the tracker against priority levels. Published utility schemes show the shape of the receiving end.
| Scheme | Priority levels and maximum response |
|---|---|
| SCE transmission programme | Priority 1: temporary or permanent corrective action within 72 hours; Priority 2: same day up to 36 months depending on condition and location; Priority 3: low-impact conditions, regulatory infractions and programmatic work |
| CPUC GO 95 Rule 18 | Level 1: act immediately; Level 2: correct within a period set by a qualified company representative, capped at 6 or 12 months for fire-risk nonconformances by fire-threat tier, 12 months for worker-safety issues and 59 months otherwise; Level 3: re-inspect, re-evaluate or repair as appropriate |
| PG&E transmission manual | Priority A immediate, B within 3 months, E within 12 months, F within 24 months |
GO 95 also lists the factors that legitimately move priority for the same condition: type of facility, location, accessibility, climate and impact on operations and the public. Those factors fit remote weather stations closely, since a mast reachable only in a summer access window may need its corrective date set differently from an identical finding at a staffed site, without its severity changing. PG&E’s priority levels are written for typical conditions and are adjusted up or down on the inspector’s judgement of what was actually observed.
In practice, a corrosion finding maps by material loss and location (surface rust on an enclosure door against section loss on a mast leg), a crack by whether it sits in a load path, and a loose fitting by whether the connection has redundancy. The mapping is organisation-specific and has to be validated against the receiving team’s actual service-level definitions; a generic table copied from another operator will misfire.
Ranking the inspection queue by asset criticality, not flag timestamp
A risk-based inspection prioritisation pipeline can sort the daily review queue by combining a per-asset criticality weight with the severity the vision model detects, so a severe finding on a single-point-of-failure radar rises above a minor flag on a redundant mast.
TechnoLynx builds this kind of pipeline through its Computer Vision R&D work: the first step is defining asset-criticality tiers and defect-severity classes with the operator’s maintenance and reliability stakeholders, then designing scoring that fuses detection confidence, defect class and criticality tier into one ranked queue, validated against expert-assessed urgency.
The criticality half has a standard method. Asset Criticality Analysis scores each asset as Consequence × Likelihood, where consequence is a weighted sum of severity scores across categories such as safety, environment, production, quality, maintenance cost and customer impact. The weights must sum to 100 percent and are set by senior leadership before scoring begins. Ranked scores are then cut into tiers, typically Critical, Important and Non-Critical. Asset criticality and risk-based inspection share the same Consequence × Likelihood framework; risk-based inspection applies it specifically to inspection programmes for static equipment with deeper damage-mechanism analysis.
Weighting a single-point-of-failure station against a redundant one
A single-point-of-failure station should carry a higher consequence score in the categories its loss actually affects, chiefly service delivery and downstream data users, while a redundant station keeps its safety consequence unchanged. PG&E’s failure modes and effects analysis was built partly to identify single points of failure by occurrence, severity and detection. Two rules from criticality practice guard against misweighting. A single-category score at the top of the scale in safety, environment or production usually warrants Tier 1 regardless of the aggregate score. Assets should be scored against their worst credible failure mode, not their most likely one.
Some conditions should bypass ordinary severity debate altogether. A collapsing mast is a safety problem whether or not a neighbouring station covers its data, which is why critical-severity findings should bypass the weighting and go straight to immediate escalation.
Today, this week or routine: an example decision matrix
| Critical | Escalate on detection | Escalate on detection | Escalate on detection |
| High | Review today | Review today | Review this week |
| Medium | Review this week | Review this week | Routine queue |
| Low | Routine queue | Routine queue | Batch with next inspection |
The matrix only stays honest if its inputs stay current. Critical assets should have their criticality reviewed quarterly, important assets annually and non-critical assets when triggered, with re-analysis after operational changes, asset additions or removals, or incidents that expose scoring gaps. For a weather network, a station losing its backup, or a new radar taking over a neighbour’s coverage, is exactly such a trigger.
Severity scores are only as reliable as the detection model on the imagery it actually receives, under the weather, resolution and occlusion conditions of the network. The ranking is decision support, not a certification of asset safety, and the top of the queue still needs a human reviewer. Any combination rule (how criticality and severity multiply or override) should be transparent and tunable by the asset owner, not buried in model code.
Who can build a risk-ranked defect triage pipeline for our observing network?
TechnoLynx engineers the vision and ranking side of defect triage for weather infrastructure: defect detection models adapted per asset type, scoring logic that combines detected severity with criticality tiers, and integration of the ranked queue into the inspection workflow. The work runs as a Computer Vision R&D engagement, starting with problem framing alongside your maintenance and reliability stakeholders, then model selection, implementation optimised for your deployment hardware and daily image volumes, validation of queue ordering against expert judgement, and monitoring once it is live.
Operators who only need chronological flagging do not need a risk pipeline at all. A team that already has a signed-off taxonomy and in-house computer vision staff may decide it only needs targeted help with validation.
Can the work-order system stay untouched if the model only sends pass or fail?
Yes: a translation layer at the integration boundary can convert the model’s continuous output into the existing binary work-order field, leaving the legacy system untouched, provided the threshold is explicitly owned, documented and signed off.
TechnoLynx would build that layer by first confirming the legacy system’s binary input contract and where translation must happen, then validating the threshold against historical labelled findings so the false-positive and false-negative trade-off is visible before anyone signs it.
Deriving the binary flag from the signed-off severity level rather than from a raw confidence score lets the cut-off inherit the taxonomy’s approval instead of becoming an unreviewed number in integration code. The threshold itself is an operational decision: set too low it floods crews with unnecessary work orders, set too high it lets defects through, and both have a cost the work-order owner answers for. Sign-off therefore belongs to the operational owner responsible for work-order prioritisation, with the engineer who owns the model documenting the trade-off.
Three caveats matter. A stale threshold degrades work-order quality silently, because the interface still looks exactly the same. Model updates can cause silent quality drift, so the sign-off has to be revisited on every model release; drift detection and cost-weighted thresholds are adjacent problems with their own methods. And compressing a score to one bit discards confidence and severity detail, which is why the full value should stay in the inspection record alongside the binary output, the same principle as keeping an original classification for auditability while a standardised value feeds reporting.
Calibration keeps the taxonomy honest after go-live
Visual severity ratings vary between inspectors, and a vision model trained on those ratings inherits the variation. The FHWA’s study of routine bridge inspections found that, on average, four to five different condition ratings were assigned to each primary element, and only 68 percent of ratings fell within one point of the average. Inspectors in that study also clustered toward the middle of the scale, avoiding extreme ratings. EPRI’s survey on wind turbine blade damage found agreement on any single category was rare, with votes usually split between two categories, and suggested treating categorisation as carrying an uncertainty of one category until concrete definitions exist.
Calibration exercises are the countermeasure: several inspectors classify the same condition independently, then compare severity, controls, due dates and the need for technical review, to find which part of the framework is unclear. For an automated pipeline, the disagreements those sessions surface also show where the training labels are least consistent. A triage system whose taxonomy has never been calibrated is ranking on a scale nobody has agreed on, however good the detection model is.
