Case study: vision system for inspecting optical fiber end faces

Dalian Optical Technology Co., Ltd., a manufacturer of optical sensors and machine-vision monitoring systems in Dalian, Liaoning, was approached by a supplier of fibre-optic patch cords with a quality-control bottleneck. Inspectors were using handheld microscopes to judge the condition and geometry of connector end faces before components entered assembly. The visual method worked, but its output depended on the operator, the microscope and the lighting available at each station.

This Case Study: Vision System for Inspecting Optical Fiber End Faces follows the project from sample definition to production commissioning. The brief required automated end-face inspection, repeatable measurement of core and cladding geometry, reliable detection of contamination and physical damage, and a permanent record linked to each ferrule. The delivered solution combined an inspection lens, controlled illumination, a fixed camera station, purpose-written image-processing software and an operator interface for exception review.

The inspection problem

Fibre-optic performance is influenced by what happens at the mating interface. Dust, dried adhesive, fingerprints, scratches or a chipped ferrule can alter the optical path and raise insertion loss or return loss. Geometry matters as well: the fibre core must sit correctly within the cladding, the end face must be clean and properly cleaved, and the ferrule must remain stable enough for repeatable connector alignment.

At the original workstation, an inspector placed each connector beneath a microscope and looked for obvious damage. Magnification made particles visible, but angled reflections and coloured coatings changed what the eye could see. Operators also had to remember several acceptance rules at once, including permitted scratch lengths, contamination limits, fibre protrusion and concentricity tolerances. Their pass or fail decision was not captured electronically.

The project therefore had to solve three separate needs: image acquisition that reveals surface and geometry defects, measurement that returns values rather than impressions, and documentation that can support traceability and customer audits. Generic object detection alone was insufficient. The system needed to distinguish a manufacturing defect from a valid coating, edge highlight, loose fibre fragment or background artefact. These requirements shaped every later optical, mechanical and software decision.

Designing the optical and mechanical station

The imaging train was built around a low-distortion lens, a suitable industrial camera and a rigid ferrule holder. The camera and lens were selected for repeatable magnification and sufficient pixel detail across the complete end-face circle. Dalian’s optical-sensor engineering experience was used to set working distance and alignment rather than relying on an adjustable consumer-style stand, where small movement could change image scale.

Illumination was treated as a measurement system in its own right. Coaxial light supports viewing of the core and surface, while oblique or directional lighting makes particles, scratches and edge damage easier to separate from the background. The illumination stage followed illumination design guidance, then adapted it to the reflective ferrule and the contrast of the connector series. A recipe-controlled light source reduced dependence on ambient factory lighting.

Mechanically, the holder centred each ferrule at a fixed focal point and constrained it against rotation. Soft contact points protected the polished end face while allowing the lens to inspect the complete aperture. The cabinet was designed for operator access, cooling and routine lens cleaning, and the camera was shielded from stray light. This combination of controlled mechanics and stable illumination gave the software repeatable inputs instead of asking it to compensate for every inconsistent setup.

Developing the inspection algorithm

The team developed a custom vision algorithm rather than treating the camera output as a simple brightness check. In its first stage, the software located the ferrule and normalised scale and rotation. It then separated the fibre core, cladding, ferrule face and background through colour, intensity and texture analysis. This normalisation made measurements comparable across different connector types, ferrule colours and manufacturing batches.

The second stage combined geometric checks with defect classification. The algorithm calculated core diameter, cladding diameter, core-to-cladding concentricity, fibre protrusion and relevant end-face angles. It also assessed surface profile across the aperture to identify chips, scratches, particles, adhesive residue and irregular bright spots. Confidence limits were more useful than a single binary result: clear nonconformances could be rejected automatically, while ambiguous images were held for an operator with the original image and measured values visible.

Validation used approved production parts, known-good samples and a training set containing deliberately introduced damage. Threshold tuning was carried out with operators who had used the old microscope process, then checked against the customer’s drawings and applicable connector inspection criteria. Every change to the recipe was versioned, and the same frozen configuration was used during acceptance testing. This made later behaviour explainable: a disputed result could be traced to the exact lens setting, lighting mode, algorithm release and sample image.

Commissioning for Australian operations

The reference installation was completed in a component laboratory in Melbourne, with quality review shared by teams in Sydney and Brisbane. It was designed for the Australian fibre-to-the-home and NBN-related patch-cord market, but also for organisations maintaining links to mining operations, data centres and undersea-cable networks. Those environments reward local traceability because a rejected connector can delay a much larger installation or field repair.

Importing the station from China required attention to Australian customs rules. The customer appointed an importer of record and checked the tariff classification for the complete machine rather than treating it as a low-value parcel. Goods valued at A$1,000 or less can generally enter under the low-value imported-goods GST process, while higher-value capital equipment normally requires a formal import declaration, applicable duty assessment and GST treatment. The commercial invoice, packing list and serial numbers were reconciled before installation to avoid a customs mismatch.

The electrical and operational documentation was also adapted for use in Australia. The supplier provided evidence relevant to electrical safety, electromagnetic compatibility and RCM requirements, while the installation complied with workplace rules at each Australian site. Spares and replacement lenses were held locally, and remote support was backed by a technician in Melbourne. Before ordering the final camera and software package, the project team compared off-the-shelf vision products with the proposed custom station; the end-face geometry and variable defect types made a dedicated build the stronger fit.

Results and operational comparison

The commissioned station replaced subjective viewing with a repeatable sequence: load the ferrule, acquire images under controlled lighting, calculate geometry, classify surface condition and record the outcome. Operators still made the final decision on ambiguous pieces, but they no longer needed to judge every apparently clean end face from memory. Production histories could be searched by serial number, and failed parts retained the images and measurements needed for root-cause analysis.

Engineering choices were also checked against broader industrial vision resources to confirm that the station followed established machine-vision practice rather than an improvised camera arrangement. The remaining work was site-specific: validating recipes against each connector family, calibrating after lens changes and training staff on exception handling. That balance—standardised hardware and software with controlled product-specific settings—was central to stable operation across shifts.

The practical benefits were faster routine inspection, fewer inconsistent manual decisions, less handling of good parts and clearer traceability. The system also gave quality engineers information that a microscope image alone could not provide, including repeatable dimensional measurements and searchable defect evidence. Results should still be demonstrated on each new ferrule type, because illumination response and defect appearance can change with coatings and connector geometry.

A practical comparison highlights the differences among the three inspection approaches:

Assessment area	Handheld microscope inspection	General-purpose vision station	Purpose-built end-face system
Illumination	Depends on the bench, lamp and operator	Usually fixed but selected for broad object detection	Controlled recipe for core, surface and defect visibility
Geometry	Often estimated by eye	Can measure basic features after calibration	Calibrated core, cladding, concentricity, protrusion and angle
Defect response	Depends on inspector experience	Detects broad anomalies and configured shapes	Classifies scratches, particles, residue, chips and artefacts
Traceability	Usually recorded manually or not recorded	Possible through image export and external software	Links images, measurements, recipe and result to each serial
Throughput	Slowest for full visual and dimensional review	Moderate, with product-specific setup work	Rapid routine checks with operator review only for exceptions
Process control	Vulnerable to technique and lighting variation	Consistent after setup, but thresholds may be generic	Versioned settings, defined acceptance limits and audit history