The Data Centric Architecture of a Factory Digital Twin
the needed skeleton crew and delivered the needed process water in trucks . According to plant personnel the FDT predictions helped the Folgers plant to be one of the first New Orleans plants to be back up and running following the hurricane 2 .
7 FUNDAMENTAL CHALLENGE / PRACTICAL LESSON LEARNED
The fundamental challenge of FDTs is that most of the underlying computational problems are NP-Complete ( Wigderson 2019 [ 15 ], Harel et al 2012 [ 14 ]). Known computational theory provides little insight as to how to proceed , but Miller et al 1991 [ 2 ] suggests an algorithm engineering approach around similar problems instances works well . Decades of large scale VirtECS FDT experience corroborates this suggestion . Essentially algorithms underlying an FDT can perform poorly in either runtime or solution quality or both .
This applies to high performance algorithms , spreadsheets , or human calculation . However , experience shows that well engineered algorithms designed for specific instance classes can perform well on similar instances . When instances differ substantially from those for which the algorithm was effectively designed , then performance can degrade unacceptably . FDT development that builds on this similar instance approach can be successful and underlying architectures must be designed to rapidly accept new learning from new instances .
A practical lesson learned from real world FDT is that developing the best RTN model is an art . An RTN model that is too detailed performs unnecessarily poorly and is expensive to develop . An RTN model that is not detailed enough is unable to accurately answer questions of interest . Consider a one task RTN model of an entire factory . Raw materials are consumed at time zero of the task and a product is available by summing step process times .
Such a one task model of a factory obviously ignores very many details and is not useful in answering very many questions involving equipment , labor and many other details . However , the other extreme of developing the most detailed model that can be reasonably conceived is often too expensive for most practical needs . Furthermore , the unnecessarily detailed approach is seductive without substantial model building experience . The best practice is to develop the simplest model that will be widely usable and only make that model more complex as practical experience dictates . This is the approach to which all FDT experts converge , though even experts sometimes add too much detail on specific projects until experience suggests how to streamline them .
8 CONCLUSION
The development and use of an FDT catalyzes process improvement in many ways . The process data needed by an FDT can be structured using the Resource Task Network ( RTN ) and translated to data structures – for example equations and variables - needed by algorithms that optimize
2 https :// www . purdue . edu / uns / html3month / 2006 / 060320 . APC . Folgers . html Journal of Innovation 119