Process outcome
Define decision, tolerance, speed and exception handling.
Knowledge · System design
Choose vision architecture that production can maintain
A vision system must work beyond the demonstration. Platform, algorithm, measurement approach, control and support must meet the process requirement through changeovers, maintenance and normal production variation.
Decision sequence
Make decisions in five steps
Platform
Choose where processing and support take place.
Algorithm
Use rules or AI according to variation and validation.
Measurement
Use 2D or 3D according to the information required.
Lifecycle
Assign maintenance, data, changes and acceptance.
Platform
Smart camera or PC-based vision?
Choose the platform by computing load, camera count, interfaces, frequency of change and available support skills. A compact system is not necessarily simpler if every change requires external support.
Smart camera
Suitable for bounded tasks, limited camera counts and local I/O. Assess tools, storage, spares and line-side support.
View smart-camera applicationsPC-based vision
Suitable for multiple cameras, heavier algorithms, central data or custom software. Include operating system, updates and replacement strategy.
View PC-based visionAlgorithm and data
Classical vision or AI?
Use fixed rules when features are stable and explainable. Use AI when product variation, texture or classification cannot be bounded reliably with fixed rules. The choice also changes validation, data management and maintenance.
Variation
Test normal parts, boundary cases, defects and future variants.
Validation
Define measurements, fault definitions and acceptance by product family.
Support
Decide who changes rules, stores data, releases models and rolls them back.
Measurement approach
Use 2D while depth does not change a process decision
2D suits presence, contour, codes, surfaces and robot guidance on a known plane. 3D is required when height, volume, tilt or spatial pose determines the decision or robot motion.
Lifecycle cost and support
Design for year two as well
Cost model
Include hardware, engineering, licences, validation, changeovers, data storage, updates, training, spares and downtime.
Ownership
Define who manages recipes, checks calibration, releases changes and supports production faults.
Practical test
Test a representative product mix including boundary cases and changeovers. Measure missed defects, false rejects, cycle time, availability and recovery time. Have production and maintenance operate the system and handle faults before releasing the architecture.
Next step
Make the architecture testable
Share the task, product variation, tolerances, cycle time, line control and available support skills. This provides the basis for a trial plan and clear system boundaries.