<![CDATA[The order of the Engineering changes (ECs) that are acted upon in complex product development programs is tightly coupled to the ordering of the component, the parameters and the activities, and this ordering significantly affects rework, cost, and lead time. This systematic review combines 24 peer-reviewed papers from four research streams: design structure matrix (DSM) and process architecture, engineering change management and propagation, deep reinforcement learning (DRL) foundations and DRL-based production scheduling. The evidence is used to build EC sequencing as a Markov decision process with the propagation indices generated from DSM as the state, change or sequencing rules as the action, and penalties for cost growth, schedule growth and triggered rework as the reward. The synthesis revealed that changes require between 1/3 and 1/2 of engineering time, 48.2% of the 41,500 change requests that were raised in an 8-year period for a sensor program were implemented, and changes were propagated no further than three engineering interfaces. In learning, a deep Q-network solved 43 of 49 benchmark games while previous methods were unable to solve the same number of games. A search-augmented network architecture beat other algorithms in 99.8% of games in the benchmark. The results show that rule-selecting deep Q-network agents achieve the best trade-off between scalability and verifiability, while search-based planners have a much greater look-ahead capability but at a significantly higher computational cost. The simulator fidelity, limited change histories, explainability and certification traceability are key challenges. A reference architecture and a research agenda are suggested for the development of the transfer of DRL from production scheduling to engineering change control.]]>
Copyrights © 2024
