Periodic Reporting for period 2 - MAS4AI (Multi-Agent Systems for Pervasive Artificial Intelligence for assisting Humans in Modular Production Environments)
Okres sprawozdawczy: 2022-04-01 do 2023-09-30
The project has identified three main AI solutions that are relevant for addressing a wide spectrum of applications in manufacturing a) Problem solving using optimization (e.g. based on search), b) Knowledge-based inferencing under uncertainty (e.g. based on semantics) and c) Machine learning (e.g. based on supervised learning).
MAS4AI aims at developing and testing a distributed and interoperable AI architecture based on multi-agents technology in such a way that it contributes to hyper-agility of European factories though modular and flexible production while at the same time keeps the humans in control of the AI technology and creating impact by spreading the technology over large groups of European manufacturing companies.
There are 5 overall project objectives:
1. Development of a Multi-Agent-System (MAS) for distributing AI components in different hierarchy layers.
2. Development of AI agents using Knowledge-based Representation with Semantic Web Technologies.
3. Development of AI Agents for hierarchical planning of production processes.
4. Development of model-based Machine Learning (ML) AI agents.
5. The system will be tested and validated upon a modular testbed pilot line and five industrially relevant large scale pilot cases (demonstrating TRL 5-6) from different industrial sectors (Bicycle industry, automotive industry, wood processing industry, bearings industry, contract manufacturing) focusing at agility and safety in manufacturing.
MAS4AI aims at developing and testing a distributed and interoperable AI architecture based on multi-agents technology in such a way that it contributes to hyper-agility of European factories though modular and flexible production while at the same time keeps the humans in control of the AI technology and creating impact by spreading the technology over large groups of European manufacturing companies.
There are 5 overall project objectives:
1. Development of a Multi-Agent-System (MAS) for distributing AI components in different hierarchy layers.
2. Development of AI agents using Knowledge-based Representation with Semantic Web Technologies.
3. Development of AI Agents for hierarchical planning of production processes.
4. Development of model-based Machine Learning (ML) AI agents.
5. The system will be tested and validated upon a modular testbed pilot line and five industrially relevant large scale pilot cases (demonstrating TRL 5-6) from different industrial sectors (Bicycle industry, automotive industry, wood processing industry, bearings industry, contract manufacturing) focusing at agility and safety in manufacturing.
In the period starting from the beginning of the project the work performed in the WP1was fully completed. The work was divided in three tasks. The first two dealt with requirements definition at two different levels. In the third task the MAS4AI system architecture was defined. The major goal of the work package WP2 is to develop a system that enables operation of multiple agents on different levels of factory’s hierarchy. The high-level ideas and the MAS4AI architecture described in the WP1 were developed using concrete technologies and software frameworks. The first two tasks, which were the main concern of the 1st project period, dealt with the system’s setup, agents’ modelling and configuration. There are three milestones in the WP2. The first milestone that dealt with modelling and setup of the MAS and was part of the 1st project period has successfully been achieved. There are two different sources of knowledge for the agent: (1) knowledge about a manufacturing application domain, which is available in the ontology. (2) Characteristics and properties individual agents, which is defined in the Asset Administration Shell (AAS). The main focus of the activities in the tasks of the WP3 in the reporting period was to define a MAS4AI ontology as well as AASs for the agents. As the result of the work in the WP3 eight draft AAS agent models and serialization by means of .aasx package files have been defined. A second outcome were a draft ontology as an RDF model serialized in a Turtle format (.ttl), several ontology APIs and instantiation in the Dydra cloud RDF store. The work packages WP4 and WP5 focused on the design and first prototypical implementation of the specific agents. As the result of the work in the WP4 and WP5 the prototypical agents for planning, optimisation and monitoring as well as coalition leader agents have been designed and developed, including standalone applications for planning and optimization meta-agents, deviation agent as well as a holonic coalition agent. They will be integrated into overall MAS4AI platform and extensively tested, validated and fine-tuned on the later stages of the project. A toolbox of the 8 optimization solutions, addressing different scheduling/ planning problems has also been implemented in the WP4 in the reporting period of the project. The major goal of the WP6 is to demonstrate and confirm the utility of MAS4AI framework in real industrial scenarios through demonstration activities. In the first project period the main activities were focused on the initial setup and implementation of the MAS4AI framework. In December 2021, the first runnable MAS4AI framework prototype of the SmartFactory-KL testbed environment was demonstrated in a form of a virtual workshop, where current results of integration and realization of the projects concepts were presented. In the first iteration we focussed on an integrated MAS4AI framework example with the resource agents for control and data acquisition of production and transport units. Furthermore, different possibilities of integration of the AI agents from the WP4 and the WP5 were accessed as well as the access to the Knowledge Base from the MAS4AI agents was discussed and tested.
All the KPIs set to track the progression towards the achievement of the projects objectives were achieved including the implementation of the core MAS4AI system, several implementations of agents and participation in standardization working groups.
All the KPIs set to track the progression towards the achievement of the projects objectives were achieved including the implementation of the core MAS4AI system, several implementations of agents and participation in standardization working groups.
There are already several results of the project that are beyond the state of the art:
1. The MAS4AI system architecture uses a novel approach to combine the holonic based ideas with the Industry 4.0 Asset Administration Shells to enable flexible decentralized control and information system and be complied to the RAMI4.0 reference architecture.
2. All the agents of the MAS4AI system are described by an Agent Service Description (ASD) which is compiled to the Asset Administrator Shell (AAS) metamodel. This creates a consistent format for the description of agents' inputs and outputs and allows interoperability across different agent based systems compatible to Industry 4.0 ecosystem.
3. AASs have been created also for the agents as the Industry 4.0 assets.
4. Knowledge modelling based on semantic web technologies is used to enhance the agents’ registry services. Agents can create complex queries to the registries knowledge bases to find the required peers and make decisions.
5. The meta-agent for planning as well as the deviation agent to track the discrepancies from the plan were defined. The meta-agent can use different multi-criteria decision-making planning/ scheduling solutions to cover all the optimization problems raised by the industrial use-cases. The deviation agent constantly tracks the execution of the plan and signals the planning agent if the intervention (replanning) is needed.
6. A novel approach of using hybrid models based on both data-driven and physics-driven methods is used to build optimization agents for all the hierarchy levels
7. Self-adaptation techniques are also used on all the levels of system hierarchy. We expect considerable improvements in productivity by using adaptive techniques.
8. The models for the agents are designed in such a way that new knowledge can be included automatically to enable self-learning systems.
1. The MAS4AI system architecture uses a novel approach to combine the holonic based ideas with the Industry 4.0 Asset Administration Shells to enable flexible decentralized control and information system and be complied to the RAMI4.0 reference architecture.
2. All the agents of the MAS4AI system are described by an Agent Service Description (ASD) which is compiled to the Asset Administrator Shell (AAS) metamodel. This creates a consistent format for the description of agents' inputs and outputs and allows interoperability across different agent based systems compatible to Industry 4.0 ecosystem.
3. AASs have been created also for the agents as the Industry 4.0 assets.
4. Knowledge modelling based on semantic web technologies is used to enhance the agents’ registry services. Agents can create complex queries to the registries knowledge bases to find the required peers and make decisions.
5. The meta-agent for planning as well as the deviation agent to track the discrepancies from the plan were defined. The meta-agent can use different multi-criteria decision-making planning/ scheduling solutions to cover all the optimization problems raised by the industrial use-cases. The deviation agent constantly tracks the execution of the plan and signals the planning agent if the intervention (replanning) is needed.
6. A novel approach of using hybrid models based on both data-driven and physics-driven methods is used to build optimization agents for all the hierarchy levels
7. Self-adaptation techniques are also used on all the levels of system hierarchy. We expect considerable improvements in productivity by using adaptive techniques.
8. The models for the agents are designed in such a way that new knowledge can be included automatically to enable self-learning systems.