Uncertainty-aware CPS design is essential when a cyber-physical system must make decisions using imperfect or delayed information. In industrial automation, sensors can be inaccurate, communication can lose or delay data, and actuators may not behave exactly as expected. Meanwhile, the physical process continues to change while software processes information in discrete steps. The result is a gap between the actual state of a plant and the state a controller or distributed application believes it is managing.
For systems that adapt their behavior, this gap can affect more than measurement quality: it can influence when the system changes, what it changes, and whether that decision supports safety, performance, or resilience. A structured design process helps engineers connect uncertainty to system requirements and choose methods suited to the application rather than treating uncertainty as a single, generic problem.
Why uncertainty matters in cyber-physical systems
A cyber-physical system combines computation with physical processes through sensing and actuation. Smart CPS add the ability to make context-dependent decisions and adapt, while collective systems coordinate across multiple devices or other entities. In a factory, energy network, building, or robot team, that coordination makes system behavior dependent on information exchanged across components as well as local observations.
Uncertainty can arise at several points in this loop. A sensor may provide noisy or missing information; network delays may make a reading stale; a component may fail; or a changing environment may invalidate an assumption. Learning and statistical methods can also introduce uncertainty into a decision. These sources can interact. For example, communication delays can contribute to inconsistent information, which may affect a decision intended to protect safety or maintain process performance.
Self-adaptation adds another design consideration: teams need to establish what triggers a change, why it is needed, where it will be implemented, what will change, who or what will perform it, and how it will occur. Without an explicit link between these questions, the uncertainty model, and system requirements, adaptation decisions can be difficult to assess.
What industrial use cases reveal about uncertainty-aware CPS design
A systematic review of European projects with industrial leadership examined demonstrators involving collective behavior, physical devices, and self-adaptation. The review selected 23 use cases from 12 projects. The examples covered areas including manufacturing, smart buildings, energy grids, robotics, emergency response, and cloud services. This scope offers a cross-domain view of design challenges, rather than a universal catalogue of every CPS application.
Across the use cases, challenges were associated with system properties that help clarify what an adaptation is intended to protect or improve:
- Safety: responding to hazards that could affect people or lead to unwanted conditions.
- Performance and price: managing resources while accounting for efficiency and cost.
- Resilience: coping with component failures, abnormal behavior, conflicts, or changing conditions.
- Cooperation and connectivity: coordinating entities and addressing communication limitations.
The properties can be coupled. Resource management, for instance, may affect both performance and cost. In robotics and emergency response, the emphasis may instead fall on cooperation and resilience, where teams must continue working toward a shared goal despite failures or uncertain conditions.
Classify the source and type before choosing a method
A useful design distinction is between the source of uncertainty and its type. Sources describe where uncertainty originates: the environment, resources, adaptation functions, goals, system model, or managed system. Types describe how it appears in operation. The reviewed use cases included network delays, missing information, measurement or communication noise, peaks in resource use, ambiguity, failures, inconsistency, and other context-related uncertainty.
The distinction matters because the same uncertainty type can arise from different sources, and a source can produce several types. Changing resources, for example, may be associated with failures, missing data, or communication limitations. Automatic learning may involve noise, ambiguity, or errors in classification. Listing both dimensions gives engineers a more useful basis for selecting a response than simply labeling a system âuncertain.â
Architecture and operating requirements also shape the choice. The computation locationâsuch as a local node, middleware, edge, or cloudâcan change which uncertainties need attention. A distributed or peer-to-peer design, for example, must account for communication delays alongside other issues. Engineers should also consider trade-offs between properties and the consequences of the uncertainty for the particular task.
Select methods by purpose, assumptions, and outputs
Methods identified across the demonstrators include data filtering and estimation, statistics and probability, reconfiguration, machine learning, verification, human involvement, and declarative programming. These are not interchangeable options. A method should be selected in relation to the uncertainty it addresses and the result the system needs.
Before selecting a method, ask:
- Which uncertainty sources and types does it address?
- Is the goal to avoid or remove uncertainty, mitigate or tolerate it, or forecast its impact?
- What inputs, assumptions, and infrastructure does it require?
- Does it produce an estimate, classification, prediction, goal or property assessment, or a request for external involvement?
- Will it run as a fixed part of the design or adapt at runtime?
For example, reconfiguration may involve adding sensors during design or changing communication interfaces during operation. Filtering and estimation can help address noisy or incomplete information. Statistical and learning approaches may support estimation, classification, or forecasting, but their assumptions and potential errors must be considered. Human involvement can also be part of the response, especially when a system should notify an operator or work with a person rather than make a decision alone.
Turn the analysis into an iterative design process
A practical workflow begins with the scenario and the componentâs intended function. Identify its processes, required properties, inputs, operating modes, and outputs. Include data from local sensors and actuators as well as information exchanged with other entities. Then identify hazardous situations and determine whether adaptation is required.
For each hazard, document the relevant uncertainty sources and types, define any thresholds that trigger a change, and select methods whose assumptions and outputs fit the task. Specify the adaptation decision and where it will be carried out. Because the system structure and available information may become clearer during design, revisit the knowledge model and repeat the assessment until the design is stable.
Conclusion: design for imperfect information
Uncertainty is a normal feature of CPS operation, not an isolated edge case. For automation teams, the practical lesson is to make it explicit: connect uncertainty sources and types to hazards, system properties, architecture, and adaptation decisions. Then evaluate candidate methods against their assumptions, inputs, outputs, and intended purpose. This disciplined approach helps engineers design self-adaptive components whose responses are considered in the context of real operating requirements.
Access Official Technical PDF