Short answer
Spare parts identification data pairs catalog-style images with part numbers, dimensions, attributes, supersessions and cross-references. Industrial and automotive distributors hold it at scale, usually in standards-based product feeds. It supports recognizing and matching parts, but studio images show no in-situ wear, occlusion or handling, and manufacturer-supplied images carry rights that need checking before any training use.
The repair bottleneck this data addresses
Many repairs stall at the parts counter, not at the machine. A technician holds a worn seal, a cracked clip or a corroded fitting and has to find the right replacement among thousands of lookalikes. The match depends on thread pitch, diameter, material, rating, mounting pattern or fitment to a model year, and the printed number may be worn off. A system that identifies parts from images and reasons over attributes could shorten that search for human technicians and for robots that fetch or fit parts.
The data to train and test such a system is catalog data: images, attributes, numbers, cross-references. It is closer to product catalog data for robot perception than to demonstration data. For tool recognition, see tool identification image data. The broader cluster is at the maintenance hub, and the humanoid data guide places this asset among the others.
What the records look like
Industrial and automotive distributors organize parts in product feeds. In the automotive aftermarket, the Auto Care Association's ACES is the standard for fitment data across light, medium and heavy duty vehicles, powersports and off-highway equipment, and PIES is the standard for product information, including attributes, descriptions, images and packaging. Both are machine-readable XML and depend on subscription reference databases (Auto Care Association). Industrial distributors use other schemas, but the shape is similar: a part record with attributes, images and relationships to other parts.
| Field group | Contents | Use in training or evaluation |
|---|---|---|
| Identity | Part number, brand, category | Label for recognition and retrieval |
| Attributes | Dimensions, thread, material, rating, color | Attribute prediction and filtering |
| Images | Studio views, angles, packaging | Visual recognition |
| Relationships | Supersession, interchange, kit membership | Cross-reference reasoning |
| Fitment | Which assets accept the part | Matching to a repair order |
| Packaging | Quantity, unit, dimensions | Handling and picking |
The difficult part of cross-reference data is that it is asserted by the catalog owner. An interchange table says part A replaces part B; whether that is true for a specific application depends on details the table may omit.
What it can support, and what it cannot
Catalog data can support: recognition of parts under clean conditions, retrieval of the right record from a photo, attribute extraction, cross-reference and fitment reasoning, and evaluation of part-matching tools. With a hand-labeled set, it can support attribute reading of markings.
It cannot support: recognition of worn, dirty, rusted or partly hidden parts as found in a machine; the appearance of parts in a bin or mid-assembly; manipulation of the part; or knowledge of how the part seats. Images, attributes and 3D models support detection, identification, pose and grasp planning; they do not show how an object behaves when handled, a distinction explained in object recognition versus manipulation data.
Catalog images also carry bias: white backgrounds, consistent lighting, and often a single best view. A model trained only on them can fail on a phone photo taken under a car. Combining catalog images with in-situ images from repair shops reduces that gap, but in-situ images are not something distributors typically hold.
Supplier types
Industrial distributors hold broad catalogs with attributes and images, usually obtained from manufacturers.
Automotive parts distributors and counters hold fitment-linked records, and counter staff interactions describing wrong parts, returns and substitutions.
Manufacturers own the underlying product data and images and often control licensing.
Parts-lookup software vendors aggregate catalogs from many sources and may impose their own terms.
For repair-shop supply that can supply in-situ images and diagnosis context, see repair shop networks as suppliers and automotive service task data. For technical drawings, see exploded views and parts catalogs.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: replacement_parts_catalog_v0
domain: industrial_fasteners_and_seals
records:
part_id: P-000417
category: o_ring
attributes: {inner_diameter_mm: 24.0, cross_section_mm: 3.0, material: nitrile, hardness_shore_a: 70}
images: [front, angled, packaging, dimension_drawing]
relationships:
supersedes: [P-000391]
interchange: [P-102233 (brand_X), P-554219 (brand_Y)]
kit_member_of: [K-0912]
fitment: [asset_class_pump_type_Q]
provenance: {feed_version: 2026-Q2, image_source: manufacturer_supplied, license_scope: reviewed}
eval_split:
held_out: by brand and by part family
query_images: phone_photos_of_worn_parts (separately collected)
Handling the long tail
Parts catalogs follow a steep long tail: a few fast-moving parts have many images and records, while most have one studio shot and sparse attributes. Report class counts openly, and decide whether you need balanced coverage or coverage that mirrors demand. For a recognition model, rare parts are usually where errors concentrate, and holding out whole families, not random images, shows how well the model handles new parts.
Look-alike clusters deserve their own test: fasteners differing only by thread pitch, seals differing by a fraction of a millimeter, and filters differing by flange. If a part can only be told apart by a measurement, no image-only model will get it right, and the benchmark should say so. Link such cases to dimension attributes and, for installed fitment, to the asset records in CMMS asset hierarchies.
Quality and acceptance criteria
- Part numbers resolve uniquely; duplicates and variants documented.
- Attributes use consistent units and controlled vocabularies.
- Images show multiple views, scale cues and legible markings where relevant.
- Supersessions dated, with direction stated (old to new).
- A query set of in-situ photos exists for evaluation, even if small, and sits outside the training pool.
- Holdout by brand and family to test generalization to unseen parts.
- Coverage statement: number of part families, brands, categories and countries of origin.
Permissions and provenance
This asset has a rights profile that differs from internal operational data. Distributors usually receive images and descriptions from manufacturers under catalog terms that may allow display and sales use but not AI training. Subscription reference databases such as those the Auto Care Association standards rely on carry their own license terms. Diligence questions: who authored the images, what license covers redistribution or machine learning, whether trademarks and logos appear, whether the feed is a licensed derivative, and whether the supplier can grant rights or must obtain them upstream. Redaction does not apply to such content; scope is a contractual matter. This is not legal advice.
Limitations and preparation
Expect work before the data is useful for perception training: deduplicate images, align them to part numbers, remove watermarks, standardize attributes and produce in-situ test images. If you want manipulation, the catalog gives dimensions but not grasp points; see tool affordances and grasp annotations for how functional labels are specified. Cross-reference tables can be wrong or outdated, so treat them as claims to validate against fitment evidence from repair records.
Questions to ask suppliers
- Which schema or standard does the feed follow, and what version?
- Who supplied the images, and what license terms attach?
- How many views per part, and are scale references or dimension drawings included?
- How are supersessions and interchanges maintained and dated?
- Do any records link to repair orders or fitment evidence?
- Are subscription reference data included, and under what terms?
- What share of parts has complete attributes?
- Can the supplier grant AI-training rights, or only display and sales rights?
Discuss your requirements
If you need parts images with attributes and cross-references, describe the part families, standards and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources. A request does not guarantee a match. The catalog entries for product catalogs and descriptions and supply chain logistics data describe neighboring asset types.
Explore sourcing and licensing options with SourceX
Sources
- Auto Care Association ACES and PIES standards — fitment and product-information standards and the reference databases they rely on.