Transmission EOL: The Wrong Baseline Costs a Whole Batch
Transmission assemblies are large and carry many inspection points, travelling on a slat line with robot transfer. These lines run frequent mixed-model production: the same station inspects one variant in the morning and another in the afternoon, and inspection content changes with the model. At that changeover frequency, the error people make most often is not missing a component but taking the wrong decision baseline.
On a mixed line the real risk is the wrong model
On a single-variant line, a missing part is the main defect mode and the eye catches it. On a mixed line, assembly completeness is only the surface. Different variants use different component combinations, so a part for variant A fitted to variant B can be in the right place and quantity, and still be wrong. That cannot be found without checking configuration, and if it reaches the vehicle the rework scope is far larger than a missing part. This cell splits verification into configuration and component presence: the first asks whether the part should exist, the second whether it is there and facing correctly.
The baseline follows the code, unit by unit
Reading the product code and assembly list determines which check set applies to this unit, and that set is the baseline. Adjacent units can differ, so the baseline must switch per unit and never carry over the previous conclusion. Code-driven switching hands the job to the system: changeover happens automatically, so nobody can forget to switch, and nobody can switch to the wrong variant.
A single switch error becomes a batch verdict
This is the danger specific to mixed lines. When the baseline is wrong, the system does not fail to decide; it decides confidently and wrongly. Apply variant A's list to variant B and missing parts are flagged one after another while extra parts are called wrong, stopping a whole batch. Code reading reliability deserves the same attention as inspection: a failed read, a misread or a read of the previous code all need explicit handling, never a default carry-over. The rule is often missed until a burst of false calls appears.
Cut the assembly area out of the frame before judging
Area cameras capture several angles, and feature extraction first separates each assembly area from the full image before checking point by point. Analysing the whole frame leaves escape risk uncontrolled: a region may look fine while the logic never actually covered it. Once the frame is divided into inspection areas, a long point list becomes structured item-by-item judgment, and it is possible to say which item was checked and which was not.
When two engines disagree, strictness is the rule
AI judges presence and pose; template matching compares precisely against the assembly baseline. They cross-check each other: agreement releases the unit, disagreement is treated strictly and routed to review. The rule looks like it lowers throughput, but it separates definite release from doubt, and a doubtful unit should not have its fate decided by an algorithm.
What 68 seconds has to contain
The cycle is 68 seconds and must match the slat conveyor and robot transfer rhythm, subject to site acceptance results. It contains code reading, multi-angle imaging, dual-engine verification and diversion. In mixed operation, variant switching cannot consume extra time, or the unit arrives while the system is still loading a template and the takt is gone. Template retrieval must finish at the instant code reading completes, so imaging and judgment can start immediately.
Trade-offs and ownership on a mixed line
Does template maintenance run out of control with many variants?
It can, and it is the main long-term cost. Variant count and template count are one to one, so every new variant needs a template and samples. Assign ownership and an update process at approval.
Do two engines still help when parts are very similar?
They help more. Similar parts differ subtly, single-method false rates rise, and two engines based on different principles cover exactly that gap.
Does 100 percent mean every point is photographed?
It means every unit's assembly configuration and component presence are visually verified, not sampled. Point coverage is set by camera layout and angle, which is outside that definition.