Code-Switched Templates Drive Mixed-Model Die-Cast EOL
Die-cast parts and their assemblies are large and structurally complex, and they travel on roller conveyors through final assembly checks for engines, new-energy powertrains, diesel engines and excavators. These lines are usually mixed-model: one unit is a diesel assembly, the next may be a different variant. Final inspection has to verify configuration and component presence item by item against an assembly list, and asking people to do that at this changeover frequency makes errors close to inevitable.
Mixed-model changeover is a template problem
Changing model means changing the decision basis: which parts belong on this variant, where each sits, which way it faces. Together that forms an inspection template. Stopping the line to switch disrupts the takt, and switching by hand drops items. This cell is code-driven: the assembly is identified on arrival, its assembly list is retrieved, and the matching template is loaded automatically, so changeover completes inside the takt without stopping the line.
Every template brings its own light
Lighting is the part easily overlooked. Different variants have different assembly-face materials and structures, so the mix of structured light and diffuse light differs. Changeover therefore switches light parameters as well as the template. Here both are issued together with the list, which prevents the common failure where the template changes and the light stays on the previous recipe; judgment then drifts for no visible reason and the algorithm gets blamed while the lamp is at fault.
The arbitration rule matters more than either engine
The AI channel judges component class and state and returns a probabilistic conclusion. The template channel checks the standard configuration item by item and returns a definite one. The two are cross-arbitrated and either engine flagging an anomaly blocks the unit. The strict rule is deliberate: a missing part that reaches a complete machine costs far more than an extra unit sent to recheck. Strictness also raises false calls, so the manual review station belongs on the anomaly branch, not on the release branch.
Reflection is a decision-criterion problem, not an imaging detail
Die-cast metal reflects strongly, so the same component presents different gray levels from different angles. If the criterion sits on absolute gray, the verdict changes with the angle. Diffuse light lifts the surface evenly and weakens specular reflection, while structured light is applied where a three-dimensional contour or edge needs emphasis, highlighting whether a part is seated and which way it faces. Once reflection is suppressed, gray consistency improves and false calls from both engines fall.
Dense points drive face-based camera placement
Area cameras at 12 to 20 MP are the usual configuration at this level, and pixel count must match point size: cover too much with one frame and the smallest part gets too few pixels; cover too little and camera count climbs. On assemblies with dense points, the imaging area is divided by assembly face, with multiple cameras where needed, so the assembly completes all imaging during one stop on the roller line and repeated positioning differences are reduced.
What fills the 75 seconds
The cycle is 75 seconds per unit and changeover happens automatically from the list without stopping the line, subject to site acceptance results. Anomalies are diverted, and inspection data is archived by code and uploaded to MES, where the image and the decision trail for each unit can be retrieved. Code reading, multi-angle imaging, dual-engine judgment and diversion fill that window; compressing it further means changing camera count or decision strategy, not adjusting a parameter.
Two preconditions for changeover and arbitration
Could we keep only one engine?
AI alone generalizes well but is less definite, and can pass subtle differences between similar parts. Template matching alone checks configuration accurately but needs a new template for every variant. They are paired because they answer different questions.
Is mixed-model changeover really stop-free?
Code reading and template retrieval are automatic and the switch completes within the takt. A new variant still needs samples and a template built first, and that work sits outside the line cycle.
What about heavily reflective die castings?
Fix the light combination and incidence angle first, then discuss algorithms. If imaging is unstable, any algorithm can only buy pass rate by loosening tolerance.