The use of machine learning solutions is becoming increasingly popular in manufacturing. Even though some approaches have proven to be very promising, there are still some hurdles to overcome when it comes to actually using them in practice. In particular, in the industrial environment, the use of machine learning methods is often motivated by the goal of reducing costs. However, conventional approaches and AutoML tools usually optimize exclusively according to a limited number of selectable metrics (esp. accuracy) without taking into account the consequences that possible errors can trigger. Cost-sensitive methods are often suggested to integrate the business view into a model, but there are hardly any usable algorithms proposed for this, especially for multiclass classification. In addition, the cost matrix is often not known, or decision preferences are anchored only as implicit knowledge in the minds of experts. According to that, we propose a framework that optimizes a model decision logic based on implicit expert preferences collected from pairwise comparison. In particular, the preference-based optimization differs from other approaches, because it does not require the costs to be explicitly defined.

The navigation of AMRs (Autonomous Mobile Robots) within factories, especially when facing moving and changing obstacles, still proves to be a challenge. ROS2 offers promising open-source behaviors that facilitate easier programming. When trying to implement reinforcement learning algorithms that aim at providing more robust policys, one is still faced with a difficult stack of technological layers. We provide an exemplary setup that shows the interplay between ROS2 (Robot Operating System 2) Humble and Reinforcement Learning. This setup works with different robot simulations, and we investigate how well a policy trained on one robot can be transferred to another.

The distribution of fuel gases is undergoing major changes due to decarbonization efforts: Non-fossil gases such as biomethane or renewable hydrogen can lead to the reuse of existing gas infrastructure for gas storage, transport, and distribution to reduce greenhouse gas emissions while maintaining a high energy security. For safe and efficient operation, we propose Gas Grid Copilot (GGC) as a demonstrator of a multi-objective reinforcement learning agent that trains in a simulated gas grid environment to control a grid by modifying its inflow into a mass storage. Multiple, possibly conflicting reward signals are included. Their conflicts and synergies of rewards are analyzed using techniques from multi-criteria decision making, more specifically a conflict interaction matrix based on extended fuzzy logic. That way, dispatchers of a gas grid can explore the effects of reward prioritizations and their consequences safely.

Machine learning models (esp. deep learning) potentially reduce costs in manufacturing by making better quality decisions or predicting machines' states accurately. The more complex the input gets (e.g., combining images and process data in automated manufacturing lines) the more data needs to be sent across one or more industrial plant's computer network to a cloud running a high capacity model. However, this approach might be less desirable due to cloud provider costs, privacy, and latency concerns for in-line quality control - or, it may even be infeasible due to security issues prohibiting a direct internet connection on the shopfloor. On the other hand, some cases might be so clear (e.g. okay part or scrap) that a weaker model running on an edge device with no internet connection can reliably perform the task. We propose a cascading architecture on the IoT-edge-cloud continuum for manufacturing use cases based on Apache Kafka that makes it possible to distinguish between easy and hard instances at every level and only delegates instances to higher capacity models if needed. We demonstrate, using an image dataset for defect detection, that this cascading approach achieves similar accuracy (about 1.5% decline) while reducing compute time on the highest-capacity model by 52%.

Carbon fiber reinforced polymers (CFRP) offer highly desirable properties such as weight-specific strength and stiffness. Liquid composite moulding (LCM) processes are prominent, economically efficient, out-of-autoclave manufacturing techniques and, in particular, resin transfer moulding (RTM), allows for a high level of automation. There, fibrous preforms are impregnated by a viscous polymer matrix in a closed mould. Impregnation quality is of crucial importance for the final part quality and is dominated by preform permeability. We propose to learn a map of permeability deviations based on a sequence of camera images acquired in flow experiments. Several ML models are investigated for this task, among which ConvLSTM networks achieve an accuracy of up to 96.56%, showing better performance than the Transformer or pure CNNs. Finally, we demonstrate that models, trained purely on simulated data, achieve qualitatively good results on real data.

Components made from carbon fiber reinforced plastics (CFRP) offer attractive stability properties for the automotive or aerospace industry despite their light weight. To automate CFRP production, resin transfer molding (RTM) based on thermoset plastics is commonly applied. However, this manufacturing process has its shortcomings in quality and costs. The project CosiMo aims for a highly automated and cost-attractive manufacturing process using cheaper thermoplastic materials. In a thermoplastic RTM (T-RTM) process, the polymerization of ε-caprolactam to polyamide 6 is investigated using an intelligent mold tooling. Multiple sensor types integrated into the mold allow for tracking of process-relevant variables, such as material flow and polymerization state. In addition to monitoring the T-RTM process, a digital twin visualizes progress and makes predictions about issues and countermeasures based on machine learning.

Resource allocation is a common problem in many technical systems. In multi-agent systems, the decentralized or regionalized solution of this problem usually requires the agents to cooperate due to their limited resources and knowledge. At the same time, if these systems are of large scale, scalability issues can be addressed by a self-organizing hierarchical system structure that enables problem decomposition and compartmentalization. In open systems, various uncertainties—introduced by the environment as well as the agents’ possibly self-interested or even malicious behavior—have to be taken into account to be able to allocate the resources according to the actual demand. In this article, we present a trust- and cooperation-based algorithm that solves a dynamic resource allocation problem in open systems of systems. To measure and deal with uncertainties imposed by the environment and the agents at runtime, the algorithm uses the social concept of trust. In a hierarchical setting, we additionally show how agents create constraint models by learning the capabilities of subordinate agents if these are not able or willing to disclose this information. Throughout the article, the creation of power plant schedules in decentralized autonomous power management systems serves as a running example.

Resource allocation is a task frequently encountered in energy management systems such as the coordination of power generators in a virtual power plant (unit commitment).Standard solutions require fixed parametrised optimisation models that the participants have to stick to without leaving room for tailored behaviour or individual preferences.We present a modelling methodology that allows organisations to specify optimisation goals independently of concrete participants and participants to craft more detailed models and state individual preferences.While considerable efforts have been spent on devising efficient control algorithms and detailed physical models in power management systems, practical aspects of unifying several heterogeneous models for optimisation have been widely ignored -a gap we aim to close.As a by-product, we give a formulation of warm and cold start-up times for power plants that improves existing power plant models.The concepts are detailed with the loaddistribution problem faced in virtual power plants and evaluated on several random instances where we observe that a significant number of soft constraints of individual actors can be satisfied if considered. I. CONSTRAINT OPTIMISATION PROBLEMS IN POWER SYSTEMSR ESOURCE allocation and scheduling are difficult prob- lems that occur frequently in energy systems, be it the coordination of power generation [1], demand-side management, or building control software.In a producer-based view, supply needs to meet the demand as accurately as possible in order to guarantee stability and avoid costs incurred by corrective measures.Similarly, consumers may try to find cost-minimising schedules for processes required throughout a day with respect to time-dependent energy prices.Current initiatives 1 are based on the assumption that groups of prosumers (i.e., energy producers and/or consumers) can form and team up to achieve better prices or production rates for their participants.We also adopt the notion of agents, indicating that the prosumers are in principle autonomous entities, even if they surrender the decision about their power output to the group.A straightforward solution (see, e.g., [2], [3], [4], [5]) to this resource allocation problem is to model the decision making process (e.g., distributing the load in a virtual power plant (VPP) or scheduling energy-consuming domestic processes in a consumer coalition) as a mathematical optimisation problem such as a mixed integer program (MIP), a linear program

Many resource allocation problems are hard to solve even with state-of-the-art constraint optimisation software upon reaching a certain scale.Our approach to deal with this increasing complexity is to employ a hierarchical "regio-central" mechanism.It requires two techniques: (1) the synthesis of several models of agents providing a certain resource into a centrally and efficiently solvable optimisation problem and (2) the creation of an abstracted version of this centralised model that reduces its complexity when passing it on to higher layers.We present algorithms to create such synthesised and abstracted models in a fully automated way and demonstrate empirically that the obtained solutions are comparable to central solutions but scale better in an example taken from energy management. 15

Partial constraint satisfaction and soft constraints enable to deal with over-constrained problems in practice. Constraint relationships have been introduced to provide a qualitative approach to specifying preferences over the constraints that should be satisfied. In contrast to quantitative approaches like weighted or fuzzy CSPs, the preferences just rely on a directed acyclic graph. The approach is particularly aimed at scenarios where soft-constraint problems stemming from several independently modeled agents have to be aggregated into one problem in a multi-agent system. Existing transformations into weighted CSP introduce unintended, additional preference decisions. We first illustrate the application of constraint relationships in a case study from energy management along with deficiencies of existing work. We then show how to embed constraint relationships into the soft constraint frameworks of partial valuation structures and further c-semi rings by means of free constructions. We finally provide a prototypical implementation of heuristics for the well-known branch-and-bound algorithm along with an empirical evaluation.