Mitsubishi Electric brings AI inspection into production with MELSOFT VIXIO

Mitsubishi Electric’s MELSOFT VIXIO is software for building and running AI-based visual inspection. It connects image preparation and model development with camera input, inspection tasks and production records. The decision is practical: whether a visual judgment that is difficult to encode with fixed rules can be made reliably enough to support a manufacturing process.

01 / ProductVisual inspection within an industrial automation portfolio

The company overview identifies Mitsubishi Electric’s industrial and infrastructure businesses. The company is distinct from other organizations carrying the Mitsubishi name. Here the relevant offering is MELSOFT VIXIO, an active factory-automation software product with separately described development and runtime licenses.

VIXIO addresses both model work and operational inspection. The workflow description presents no-code task construction, camera and PLC communication, monitoring and traceability. Inspection images and results can be associated with production information. That connection matters when a quality team needs to investigate which item was inspected, what the system saw and what action followed.

The product catalog describes anomaly detection using normal images and other inspection methods for defects. A model that identifies deviations is not automatically a classifier of every possible failure. Its output needs to be interpreted against the manufacturing specification: a visible difference may be harmless, while a consequential defect may be difficult to see in the chosen image.

02 / AudienceFor production teams that can define an inspection decision

VIXIO is relevant to manufacturers with a specific visual quality problem and an engineering owner for the station. Examples might include unusual surface appearance or variation that makes a fixed rule brittle. The first question is not how many model features the software offers. It is what constitutes an unacceptable item and how reliably that condition appears in the camera view.

The software is less likely to solve a problem dominated by poor lighting, unstable fixturing or an invisible defect. A process may need better image acquisition or another sensing method before AI adds value. A useful feasibility stage separates those physical limitations from limitations of the learning method, rather than repeatedly retraining on inconsistent images.

The Cognex blueprint offers a comparison for integrated machine-vision systems, while the Keyence blueprint is relevant to buyers comparing inspection hardware and deployment support. VIXIO should be evaluated as part of a station design. Compare the complete route to a maintainable inspection result, including who owns cameras, controllers and production acceptance.

03 / WorkflowA proposed inspection station for variable surface defects

Consider a proposed station inspecting the appearance of molded housings. This is an illustrative workflow, not a tested Mitsubishi Electric installation. Start with a written defect taxonomy and representative examples agreed by quality staff. Record which decisions can be automated and which require secondary review. A disagreement between human inspectors must be resolved before it becomes contradictory training data.

Control the imaging setup next. Define lighting, exposure, lens, positioning and the inspected area. Collect images across shifts, production lots and acceptable cosmetic variation. Keep a separate evaluation set, including the difficult borderline cases. Avoid allowing near-duplicate images of the same item to appear on both sides of a development and evaluation split.

For an anomaly-detection approach, use normal examples that cover acceptable variation rather than only pristine demonstration parts. Review where the model highlights differences and determine whether those regions correspond to meaningful quality decisions. For known defect classes, establish the available evidence for each class. A tool’s ability to start training from few examples does not establish production accuracy.

Build the inspection task around the actual sequence of events: trigger, capture, analysis, decision and downstream handling. The VIXIO specifications identify supported camera standards and MELSEC controller families, with a separate usable-camera bulletin. Confirm the exact camera and controller combination. Reading images from a folder is a different integration path from controlling a camera directly.

Use timestamps and production identifiers to connect an image to the correct housing. Define behavior for a missing trigger, delayed image, unavailable controller and uncertain result. An inspection station must not report a successful pass merely because no result arrived. The quality and controls teams should agree how an incomplete inspection affects the physical item and the production record.

The inspection examples include combining conventional rules with AI and using human secondary inspection. Those are useful deployment patterns, not evidence that one pattern always improves accuracy. In the housing example, a deterministic dimension check might remain separate while AI flags unusual surface appearance. Keep each decision’s reason visible to the person investigating a rejected item.

Run an agreed pilot against independent inspection records under representative conditions. Compare false accepts, false rejects, cycle time and the workload passed to people. Set thresholds with the cost of each error in mind. These are proposed evaluation measures; Sequenced has not performed a VIXIO benchmark or verified a defect-detection rate for this scenario.

04 / PricingDevelopment and runtime are separate purchasing decisions

The product page describes node-locked licenses for one computer. A development workstation and several production stations therefore need a deliberate licensing plan. Ask how a failed computer is replaced, how licenses are transferred where permitted and whether a spare station needs its own entitlement. A perpetual software license is not the same thing as an indefinite support commitment.

The public pages do not provide a universal numerical price. Request a quotation with product codes, region, taxes, support scope and any training or integration services. The catalog describes a temporary license associated with the installer; confirm its current conditions before relying on it for an evaluation. Do not equate a trial route with unrestricted production use.

Budget the physical inspection station separately. The cost of fixturing, lighting and controls work may exceed the software line item for a difficult application. Include the quality team’s time to assemble evidence and investigate errors. A low software price would not compensate for a system that shifts excessive false rejects to manual reinspection or loses traceability.

05 / DistinctionsTraceability makes the model part of a factory process

The distinctive value is the bridge between model creation and production context. A classifier output becomes more useful when the operator can recover the associated image, item and process information. That can shorten investigation and make a threshold change understandable. The organization still has to define record retention and the authority to change an inspection recipe.

Another useful distinction is the combination of AI and conventional automation. A model can address appearance variation while established controls continue handling timing and machine behavior. Keeping these responsibilities explicit makes the system easier to diagnose. If a part is rejected, the team should know whether the cause was a visual judgment, missing data or a separate process interlock.

06 / QuestionsSupported hardware and operating conditions need an exact answer

The specifications distinguish hardware requirements by task: anomaly detection alone does not require an NVIDIA GPU, while other AI tasks have GPU requirements. Confirm the current supported configuration and chosen model workload before purchasing the PC. A general statement that VIXIO can run without a GPU would conceal a consequential restriction.

Camera compatibility deserves the same precision. An interface standard is not blanket approval of every device that mentions that standard. Check the listed usable cameras and required settings, then validate the complete acquisition path. Network topology, synchronization and camera placement belong in the station design even when the software can connect successfully on a bench.

Finally, plan for changing production. New materials, surface finishes and lighting maintenance can change the image distribution. Define when the model must be reevaluated and retain the previous accepted configuration. Production evidence should determine whether a revision is better. A new model version or a lower training error alone does not establish an improvement on real items.

07 / DecisionSelect VIXIO around a measurable quality decision

VIXIO is worth evaluating when a factory needs AI visual judgments connected to camera tasks and production records. Start with one inspectable defect family and a defined route for uncertain cases. The strongest purchase decision rests on representative evidence, a maintainable station design and explicit responsibilities for model changes and quality acceptance.