A multi-agent MES can give manufacturing execution systems a way to connect shop-floor data with timely operational decisions. In a proposed smart factory architecture, software agents represent production assets, collect process data, and work with a central MES-based agent to identify unusual behavior. The concept illustrates how machine learning in manufacturing could extend MES beyond monitoring and recordkeeping toward coordinated decision support.
Why extend MES with decision support?
Manufacturing execution systems sit between enterprise-level planning and physical production. They support production planning and control while providing access to shop-floor information. As connected equipment generates increasing volumes of data, however, simply collecting and storing it is not enough. Operations teams also need ways to find relevant patterns, relationships, and potential problems in time to act.
Machine learning (ML) offers tools for analyzing production data and identifying behavior that may not be obvious from individual readings. Its value in this context is not limited to optimizing a single machine or task. The broader opportunity is to combine analytics with existing manufacturing information systems so that insights can inform production coordination. This systems perspective brings software, equipment, processes, and decision-making together.
How the multi-agent MES architecture works
The proposed architecture uses a centralized multi-agent arrangement. A sub-agent acts as a virtual representation, or digital shell, of an individual physical asset. In the concept, sub-agents run on Raspberry Pi platforms, monitor production, and collect data from the equipment. A central agent runs as middleware on the MES server and performs the analysis and coordination.
The design follows a client-server approach. Asset-level agents provide production information to the central agent, while the central agent can return commands to the relevant sub-agent. A repository on the server holds information about system topology, helping the central agent identify which asset or agents are associated with a detected issue.
The proposed workflow consists of several connected stages:
- Monitoring and collection: Each sub-agent observes its asset during production and stores the data it collects.
- Model building: Selected production data is sent to the central agent to establish a model of normal behavior.
- Anomaly detection: The central agent evaluates new data against that model.
- Coordination: If behavior is classified as abnormal, the central agent identifies the associated sub-agent or agents and determines a response.
- Behavior adjustment: Commands are sent through the relevant sub-agent to affect the physical asset. In the example, the proposed action is to switch off the machine.
MES anomaly detection using production data
The example focuses on two potential anomalies on a production line: drilling speeds that are unusually high or low, and an unusual number of parts completed per minute. These variables—drilling speed and parts per minute—are the selected features for the detection model.
The approach assumes that normal production data for these features follows a Gaussian, or normal, distribution. Using sample data, the model estimates the mean and covariance, which describe the data’s central tendency and variation. For each new observation, the central agent calculates its probability density under the model. If that probability falls below a predefined threshold, the observation is treated as an anomaly.
This is a statistical method for distinguishing data that is consistent with the modeled normal behavior from data that is less likely under that model. The threshold is a decision parameter: it determines how unusual an observation must be before the system flags it. The proposal does not establish a universally appropriate threshold or claim that the method is suitable for every process. In an industrial application, the features, training data, and detection criteria would need to reflect the production context.
From detection to coordinated action
A key feature of the concept is the link between analytics and execution. An anomaly flag is intended to initiate a decision process, rather than remain an isolated alert. The central agent consults the system topology, associates the event with the relevant asset, and sends an adjustment command through its sub-agent. For the scenario described, that adjustment is turning the machine off after abnormal behavior is detected.
That action is presented as an example of autonomous production execution, not as a proven safe response for all anomalies or equipment. The proposal outlines an architecture and algorithm; it does not report a completed industrial validation or demonstrate that automatic shutdown is appropriate in every operating situation. Any implementation would need to define how detection results, equipment behavior, and operational responsibilities fit together.
Practical considerations for manufacturers
For engineers and system integrators, the concept highlights several design questions that arise when adding ML to MES:
- Data selection: Which process variables provide a useful representation of normal and abnormal behavior?
- Model assumptions: Does the chosen statistical model adequately represent the production data?
- System integration: How will asset agents exchange data and commands with MES, and how will the topology be maintained?
- Response logic: What should happen when an anomaly is detected, and which assets or processes are involved?
- Operational oversight: How should automated decisions relate to existing production controls and human responsibilities?
The proposed work identifies communication between MES and agents running on programmable logic controllers or Raspberry Pi platforms as an area for further development. It also describes an open MES framework effort intended to support factory-floor interoperability through vertical integration. These points underline that analytics is only one part of the solution: reliable data exchange and a defined operational response are also central to the architecture.
Conclusion: a path toward more intelligent MES
A multi-agent MES combines asset-level data collection with centralized analysis and coordinated commands. The example uses drilling speed and production rate to demonstrate how a Gaussian-based anomaly detection method could identify unusual behavior and trigger a machine shutdown response. Its broader contribution is a design concept for embedding AI-based decision support into manufacturing execution systems.
For manufacturers, the practical implication is to consider analytics, agents, MES integration, and operational decision rules as parts of one system. The architecture remains a proposal requiring implementation and evaluation, but it offers a structured way to think about how shop-floor data might support more responsive production coordination.
Access Official Technical PDF