Siemens prepares circuit-board inspection before the first prototype

An inspection system can be waiting when the first circuit board arrives. Siemens and NVIDIA showed that approach at Erlangen on 19 October. Process Preparation X uses digital twins and Cosmos models to generate training images of acceptable boards and possible defects before a physical prototype exists.[1][2]

Siemens reports up to 85% less engineering effort and up to 40% shorter time-to-market from virtual preparation. Its event materials give no test population or detailed baseline behind those maxima.[1][2]

Build variation into the training set

Synthetic data gives the designer control over the examples a model sees. NVIDIA’s general Replicator guidance describes varying an object’s position, appearance, lighting and background within a simulated scene. The aim is to make a model learn useful visual features across conditions, rather than depend on one particular rendering.[3]

That guidance also separates differences in appearance, such as reflections and material rendering, from differences in scene content, such as which objects appear and where they sit. These explain why a convincing simulated image can still leave gaps in what an inspection model has learned. This is general training guidance, not a disclosed specification of the Erlangen implementation.[3]

For an imagined board-inspection programme, the engineering team could define both a missing component and the acceptable variation around correctly fitted components. Its test would ask whether the inspector distinguishes the two across the lighting and viewing conditions it will encounter. Making more pictures of the easy case would add less value than covering a difficult distinction.

The saving can come from the calendar

Consider an invented schedule with three tasks: four weeks to obtain a prototype, two weeks to prepare inspection, and one week to validate the process. Assume preparation can start immediately from an agreed digital design, adequate staff are available, and the final validation still needs the physical board.

The programme finishes two weeks earlier, a reduction of about 29%, without assuming any reduction in engineering labour. The calculation is 7 weeks versus 5 weeks: the overlapping stage takes the longer of four and two weeks, followed by validation. If another essential task takes eight weeks, that task can become the constraint instead.

This is why a factory needs to know where preparation sits in its schedule. Removing a delay on the longest chain of dependent tasks can bring a launch forward. Accelerating a task that already finishes comfortably ahead of the next stage may mainly release staff capacity. Both can be valuable, but they enter a business case differently.

Prepare the next production decision too

The Erlangen event also showed agents proposing production plans for human selection, with plans checked in a physics-based factory twin before implementation.[2]

A useful planning comparison would hold the customer’s required delivery fixed, then ask which arrangement uses the available people, machines and materials best. The model needs an operational choice to resolve: for example, whether to change a line’s sequence now or absorb more work on another line. A simulated plan becomes commercially useful when it exposes that trade-off before the change consumes factory time.

Earlier preparation creates room to discover and resolve problems while changes are still easier to make. The prototype and the production line can then test decisions that engineers have already had time to develop.

Report an error or suggest a correction