How far can photo-based spare-part identification go in a field-service app?
A field-service app should treat image recognition as a candidate generator, not as the spare-parts master or an automatic purchasing authority. Start with the asset, serial number and configuration; use the photograph to rank a short list; apply compatibility and supersession rules; then ask a technician or parts specialist to confirm the result. The confirmed part can create a case, quotation or order draft. Removing that final check is appropriate only when a barcode, nameplate or other deterministic evidence closes the identity gap.
The investment case is workflow, not a camera demo
Technicians often meet installed, worn or unpackaged parts whose numbers are hard to read. Visual search can reduce time spent describing the item or paging through a catalogue. It does not, by itself, know that two identical housings contain different voltages, connectors or firmware; that an old number has been superseded; or that a part is unavailable in a region.
TRUMPF provides a useful production example in its official object-recognition workflow. A user photographs a part, the Service app suggests a material number, the user selects the correct item and then submits a spare-parts clarification case. TRUMPF says the underlying database is continually expanded. It also lists practical conditions such as a mobile device, connectivity and a MyTRUMPF account, while direct ordering depends on country support for its E-Shop. The important pattern is not “AI recognises everything.” It is image assistance embedded in an accountable service and commerce journey.
Google Cloud Vision Product Search describes the generic retrieval mechanism: a product can contain reference images from several viewpoints, and a query image returns a ranked list of visually and semantically similar products. That supports a short-list design. It does not establish that a generic visual service understands a manufacturer's serial-number ranges, bills of material, approved substitutions or stock policy.
Choose the simplest reliable identification route
| Evidence available in the field | Preferred route | Product decision |
|---|---|---|
| Readable barcode, QR code or part number | Scan or OCR | Use the deterministic identifier first |
| Installed or worn part without packaging | Visual candidate search | Show alternatives and require confirmation |
| Near-identical parts across equipment revisions | Asset identity plus BOM/compatibility rules | Do not let appearance decide |
| Technician needs an expert opinion rather than an order | Photo-linked service case | Preserve context and route to a specialist |
| High-volume, standard parts with recoverable ordering errors | Combined recognition and an order draft | Automate data entry before automating commitment |
If search volume is low or the catalogue lacks stable identifiers and approved images, improving labels, scanning and the support lookup may create more value than training a visual model. A successful demonstration on selected photographs is not yet a business case.
Design a six-stage controlled journey
1. Establish the asset context
Open the function from a registered asset, serial number, work order or scan. Retrieve the model, delivered configuration, installed options, customer access and service status before searching. When the asset is unknown, the interface should make that uncertainty explicit rather than silently widening the catalogue.
2. Capture evidence that separates candidates
Guide the user on angle, distance and lighting, and request a second image of the nameplate, connector, mounting pattern or installed position where necessary. Record the image against the case and asset. Define retention and training consent because a field photograph may contain customer premises, people, serial numbers or unrelated documents. Reject low-quality and multi-object images instead of returning false precision.
3. Present a ranked short list
For each candidate, show the part number, name, distinguishing feature, reference image and confidence information. A score reflects the model and threshold used on that input; it is not a probability of mechanical or electrical compatibility. Some families may permit a top-three list, while visually ambiguous or safety-relevant families should go directly to a specialist.
4. Apply the authoritative business rules
Check model, serial range, BOM revision, handedness, dimensions, electrical rating, market, lifecycle state and supersession chain in the parts system. Vision answers “what looks similar”; governed master data answers “what may be installed and sold.” A disagreement stops the automated path and explains what needs review.
5. Make human confirmation useful
The technician or parts specialist can confirm, select another candidate or mark the image inconclusive, with a structured reason. Retain the proposed list, final decision and relevant evidence to expose weak photography guidance, incomplete master data and difficult classes. Do not feed every correction straight into training without approval and quality review.
6. Draft the transaction before committing it
An initial release should create a service case, request for quotation or cart draft. The user still confirms quantity, price, lead time, delivery address and any return-core requirement. Reduce confirmation only after identity, compatibility, availability and entitlement all come from verifiable sources and an erroneous order can be intercepted or reversed.
The feature belongs in the same service data model as service requests, spares and maintenance contracts. A stand-alone label detector leaves the hardest operational work unresolved.
What the client must prepare
The project ceiling is determined by four assets:
- Part master: stable number, description, specification, unit, status and supersession.
- Equipment relationship: model, serial range, BOM or configuration version, approved and prohibited combinations.
- Reference imagery: several viewpoints, installed and worn conditions, realistic lighting and confusing negatives, with provenance and usage rights.
- Operational outcomes: expert selections, returns, wrong shipments and reasons for an inconclusive result.
Split training and test data by time, equipment or acquisition batch so near-duplicate photographs of one part do not appear on both sides and inflate performance. New products, packaging and substitutions need versioned updates. A retired item must leave search, caches and recommendations consistently.
Acceptance needs consequence-aware metrics
AWS's Custom Labels evaluation guidance separates true positives, false positives and false negatives and reports precision, recall and F1 by label. It also explains the common threshold trade-off: raising a confidence threshold tends to increase precision while reducing recall. A spare-parts product must choose that trade-off from the cost of a wrong suggestion, not from the service default.
Measure at least the following by part family, equipment model and capture condition:
- top-one accuracy and top-three hit rate;
- incompatible parts incorrectly presented as usable;
- no-result, retake and specialist-escalation rates;
- time from capture to confirmed candidate and case or order draft;
- human overrides, cancellations, wrong shipments and return reasons;
- time for a new part to become searchable and for a retired part to disappear.
For a component whose misuse can damage equipment or create a safety risk, minimise false positives and escalate uncertainty. For a support tool whose purpose is merely to prepare a short list, top-three coverage and saved review time may matter more. Production thresholds must come from representative client samples; a cloud-service example cannot be reused as a launch commitment.
Common procurement mistakes
Treating high confidence as permission to order. Visual confidence contains no evidence about compatibility, stock, price or the buyer's authority.
Assuming more images automatically improve the model. Repeated angles, weak labels and copied catalogue shots reproduce the same blind spots. New data should add genuine variation and hard negatives.
Sending every failure to support without structured reasons. The organisation then cannot tell whether to improve capture guidance, catalogue data, class boundaries or thresholds.
Building recognition before integration. Without a stable part number, compatibility model and supersession chain, the result cannot safely enter quotation, inventory or ordering.
Seven decisions before approval
- Which parts are genuinely difficult to identify, and what does the current lookup consume?
- Can scanning, OCR or the asset serial number solve most of the problem first?
- Which visually similar parts must never be confused, and what is the consequence?
- Which system owns parts, BOMs, compatibility, substitutions and availability?
- Who owns the photographs and may they be retained for testing or training?
- Does the output create a candidate, case, quotation or order draft, and who confirms it?
- Who maintains new, retired and superseded items and re-runs acceptance tests?
The useful outcome is not an app that appears more intelligent. It is a shorter field workflow that still preserves part identity and purchasing accountability. A first release that reliably links asset context, ranked candidates, human confirmation and the service record is usually more valuable—and more testable—than a promise of fully automatic ordering.
Sources
- TRUMPF: Object recognition, covering photo-based material-number suggestions, user confirmation, service cases and operating conditions; accessed 10 October 2026.
- Google Cloud: Vision API Product Search documentation, covering reference images, multiple viewpoints and ranked visual retrieval; accessed 10 October 2026.
- AWS: Metrics for evaluating your model, covering per-label precision, recall, F1 and threshold trade-offs; accessed 10 October 2026.