COGNIPLANT: An innovative vision of the digital twin

The COGNIPLANT project presents an innovative vision of the digital twin in process industries. To this end, an advanced digital twin is being developed, building in a semantic enterprise-scale knowledge graph, and based on the knowledge obtained at the Advanced data analytics layer, named Co-Analyse.

The innovative vision of the digital twin will offer a logical model based on how information flows throughout the life of the plant, in contrast with the traditional heuristic models of the physical plant. The digital twin will capture the physical processes of the industrial plant, also incorporating the real actions, by considering rules, events, sequences, patterns and models, in a natural and direct way. Thus, the digital twin will allow the users to measure process efficiency without the need for parameterisation or modification of the schemes modelled.

A digital twin will be created for each one of the project’s demo cases, namely mining, chemical, steel and quicklime sector. This will enable the simulation and optimisation of these industrial processes according to the KPIs proposed for each scenario. These digital twins will provide a decision support system for the COGNIPLANT industrial scenarios, which will be adapted to the different roles involved in the production activities.

The digital twins will collect the necessary information from the data lake of the cognitive platform and then, based on different machine learning models (predictive, prescriptive and production optimisation developed) and the characteristics of production processes, it will offer an interaction that enables simulation and optimisation of a given KPI.

The digital twins will be used to optimise and improve the different KPIs. These optimisations will depend on the use cases, being able, for example, to reduce defects in foundry processes, to optimise maintenance, to improve costs and/or quality, productivity, etc. The digital twins will allow the user to interact with the models, allowing them to set the boundaries that constrain the optimisation (e.g., limits of machines, operators, policy on product quality, costs, etc.). These restrictions can also be given by the production process itself. Also, the user will choose which variables will be used as decision variables in the optimisation problem, to improve the KPIs.

The digital twins will also be used for simulation purposes, allowing users to change the values of some input variables and study how these changes affect the output variables of the models. This enables the study of the relationship between input and output variables of the different models. Therefore, the simulations will allow the impact of each input variable on the KPIs to be studied to provide an understanding of how input variables can be changed to improve the processes.

This approach will answer specific questions for experts about what would happen if certain parameters of any node were modified in fictitious or simulated scenarios. Also, the approach lets the user understand how the knowledge of these modifications is inferred and “exploded” throughout the network, and which indicators in other nodes are affected. Also, a “visual analytics” module will be developed to explain the decisions made (e.g. visualisation of KPIs that guide the decision support mechanism).

Cognition refers to all processes that transform, reduce, elaborate, store, recover, and use input data, to solve Industry 4.0 use cases, i.e. condition monitoring, anomaly detection, optimisation, and predictive maintenance. Cognition plays a key role in decision-making for complex, ill-structured situations.

Cognition in COGNIPLANT is based on a DSS. While DSS are quite common in the manufacturing industry, where processes are well defined and fixed, the application of advanced DSS systems in process industries still represents a challenge, which is the core of the innovation proposed by the COGNIPLANT project.

In addition, reactive programming is totally innovative in these industries, as it requires continuous updating of machine learning models, with the consequent challenge related to the computational ability to do it online. This means that nowadays this type of planning is hardly found in industrial plants.

In order to tackle this challenge, a novel method for network topology design will be proposed. This method will formulate the topology optimisation as a graph optimisation problem, by capturing both the objective and constraints in graph invariants, from the system kernel of the process mining topology, to the edge level (edge computing). COGNIPLANT’s graph theoretic approach leverages new studied mathematical techniques to provide a more systematic alternative to traditional approaches for network design. Specifically, optimisation will be done over some known graph invariants, such as maximum node degree, topology diameter, average distance, and edge betweenness, as well as over a new invariant called node Wiener impact.

In order to enable online control and optimisation, COGNIPLANT will develop deep (reinforcement) learning services that robustly predict, e.g., anomalies and trends in online and historic industrial processes, as well as viable multi-scale spatial-temporal decision models for the distributed optimisation and control of such processes. These services will provide critical insights in cause-effect relationships between product quality and process sustainability / performance. Insights will be provided also on the actually applied and the virtual (agent-based) decisions on distributed process optimisation and control paths / schemes under diverse process conditions and rages of uncertainty levels. The project team has experience in the derivation of agent based (distributed) reinforcement learning techniques for optimisation of real-time processes.