Knowledge-Graph-Enabled Product Digital Twin for State-Conditioned Recovery in Resilient Manufacturing-as-a-Service

Operational resilience in Manufacturing-as-a-Service (MaaS) networks depends on maintaining and restoring order fulfilment when disruptions affect manufacturing services whose capacity is shared across multiple products. A disruption can extend beyond the initially affected order because recovery alternatives may compete with existing capacity commitments. Repairing only the disrupted order may therefore overlook spillover effects, whereas reoptimising the entire network may unnecessarily reopen unaffected commitments and increase computational effort. The recovery boundary therefore depends on the capacity and reservation state of the shared services at the time of the disruption, which current Product Digital Twin and manufacturing knowledge-graph representations do not derive. To address this limitation, this study proposes a knowledge-graph-enabled Product Digital Twin (PDT) that derives the cross-order operational context relevant to a specific disruption. The Slack-Absorbing Causal Cascade (SACC) derives an event-specific recovery subgraph by determining which product, operation, service, and reservation dependencies become operationally relevant under the current network state. Potential dependencies are represented through process precedence, admissible substitution, and shared-capacity reservation relations, while their activation depends on the current capacity and reservation state. A disruption is expressed as a capacity-equivalent deficit; available slack absorbs the impact locally, and only the unresolved deficit activates and propagates through these relations, including to other orders whose reservations may be displaced. The resulting event-specific recovery subgraph defines the decision boundary of a risk-aware multi-order assignment and scheduling problem, while commitments outside the subgraph remain fixed. Across 200 simulated disruptions, the proposed method reduced the recovery-model scope by 44% and median decision time by 52% relative to full-network risk-aware reoptimisation, while reducing spillover delay by 21% relative to direct-impact repair. Retained throughput remained close to full-network recovery, 94.1% versus 94.6%, while on-time delivery reached 92.8% against 93.5% and the median number of changed assignments decreased from 52 to 17. For the literature, the results show that the recovery boundary can be derived from the event-time network state instead of being fixed in advance, extending PDT and knowledge-graph reasoning from semantic reachability to state-conditioned recovery. For industrial practice, MaaS operators can confine the disruption response to the commitments an event actually affects, reducing plan churn and provider renegotiation while retaining network-level performance.