Abstract Reliable non-destructive hardness measurement, such as using magnetic properties, is a critical requirement in industrial quality assurance and structural health monitoring, where destructive testing is impractical or economically infeasible. Despite its importance, extracting physically meaningful and predictive features from frequency-domain magnetic signals remains challenging, particularly in scenarios where annotated measurement data are scarce and expensive to obtain. In this study, the challenge of extracting physically meaningful features from limited data is addressed by proposing a complex-valued autoencoder. This architecture is designed to jointly encode magnitude and frequency information into a compact latent representation, thereby preserving phase-sensitive structural dependencies that are typically lost in standard approaches. The method was evaluated on a real-world dataset of high-strength alloys with sparse annotations. It is demonstrated that the proposed complex-valued representations yield significantly improved regression performance, characterized by lower prediction error and reduced variance compared to a classical real-valued baseline. These results indicate that leveraging the algebraic structure of complex-valued networks offers a superior approach for non-destructive testing in low-data regimes, enabling more stable and accurate industrial inspection systems.

Ensuring the stability and robustness of multimodal autoencoders is critical for their optimization and deployment in safety-critical industrial environments. This paper presents a rigorous analysis of Lipschitz properties in multimodal fusion, identifying theoretical vulnerabilities in standard summation and concatenation strategies. We derive explicit Lipschitz bounds for these methods, demonstrating their susceptibility to gradient instability and noise propagation as the number of modalities increases. To address this, we introduce a Lipschitz-regularized attention-based fusion mechanism that explicitly bounds gradient sensitivity through spectral normalization and dimension scaling. Empirical validation on four industrial robotic datasets, including the real-world RoboMNIST dataset, confirms our theoretical findings. On the RoboMNIST dataset, our method achieves an improvement of 54.7% in bimodal and 57.4% in trimodal reconstruction over standard attention, alongside a 20.5% gain in subordinated fault detection, while effectively regularizing the Lipschitz of the gradients.

In industrial applications, variations in system configurations, degradation modes, and the lack of labeled, run-to-failure data contribute to poor generalization by a domain-specific network on data from a different domain, although for a similar task. Adversarial domain adaptation attempts to mitigate this by introducing a domain classifier that discriminates between the latent features of source and target domain encoders, thereby encouraging the latter to learn domain-invariant representations. However, the discriminator’s susceptibility to oscillatory and exploding gradients negates the convergence of the target network. We present a novel Gradient Regularized Adversarial Domain Adaptation (GRADA) framework that utilizes a parameter-aware Adaptive Gradient Shrinkage (AGS) operator to regularize the discriminator. The discriminator parameters and gradients are used to develop a criterion matrix and a subsequent adaptive threshold, through which soft-thresholding is performed to suppress noisy gradients and preserve informative ones along their actual descent paths. Theoretical proofs validate the non-explosive and stabilizing properties of the AGS operator. We also incorporate AGS into an existing contrastive adversarial framework to test its use as a regularizer irrespective of the underlying framework. Our experiments focus on the remaining useful life (RUL) estimation task using the C-MAPSS dataset, as it provides suitable cross-domain scenarios with variable operational settings and fault modes.

Neural network architectures designed for function parameterization, such as the Bag-of-Functions (BoF) framework, bridge the gap between the expressivity of deep learning and the interpretability of classical signal processing. However, these models are inherently sensitive to parameter initialization, as traditional data-agnostic schemes fail to capture the structural properties of the target signals, often leading to suboptimal convergence. In this work, we propose a prior-informed design strategy that leverages the intrinsic spectral and temporal structure of the data to guide both network initialization and architectural configuration for the Bag-of-Functions. A principled methodology is introduced that uses the Fast Fourier Transform to extract dominant seasonal priors, informing model depth and initial states, and a residual-based regression approach to parameterize trend components. Crucially, this structural alignment enables a substantial reduction in encoder dimensionality without compromising reconstruction fidelity. A supporting theoretical analysis provides guidance on trend estimation under finite-sample regimes. Extensive experiments demonstrate the effectiveness of this approach: the informed initialization reduces reconstruction error by 15.1% on complex real-world benchmarks. Furthermore, the trend-informed dimensionality reduction yields up to a 30% decrease in model parameters and computational complexity, accelerating inference and stabilizing optimization without altering the core training procedure. • Data-driven spectral priors guide neural network initialization and depth selection. • Trend-informed initialization enables compact architectures without accuracy loss. • Informed priors stabilize optimization and accelerate convergence across datasets. • Improved efficiency achieved with reduced parameters and computational cost.

• Rigorous analysis proving the marginal convexity for multimodal autoencoders. • Block-wise optimization strategy for multimodal architectures. • Novel Lipschitz-based adaptive learning rate for scale mismatch. • Evaluation on real-world heterogeneous multimodal datasets.

This article introduces a novel integration of deep model-based learning with modular state-based Stackelberg games (Mod-SbSG) for distributed self-optimization in manufacturing systems, using a sample-efficient approach. Model-free Mod-SbSG requires frequent interactions with real systems to find optimal solutions, which can be costly, time-consuming, and risky in industrial settings. Prior studies handled this by using digital representations to train Mod-SbSG players, but accurate representations are often difficult to develop. Hence, our framework replaces digital representations with deep learning methods that learn system dynamics, optimize policies within Mod-SbSG, and reduce real-world interactions. The method includes two main steps: 1) designing deep learning models to predict system dynamics and 2) training Mod-SbSG players in virtual environments. We evaluate single- and multistep predictors and demonstrate network reuse for transfer learning in adaptable systems, which reduces real system interactions by 77.78% in a laboratory testbed industrial control scenario.

Multimodal learning is critical for robust perception in complex systems, yet integrating diverse modalities while ensuring fault tolerance remains a significant challenge. This paper presents a novel approach for fusing multimodal latent representations using a denoising variational autoencoder framework, where the fusion is achieved through the multiplication of probability density functions corresponding to each modality. By modeling each modality as a Gaussian distribution in the latent space, we derive a fused representation that optimally combines information from all modalities while preserving their probabilistic structure. We introduce a failure injection mechanism during training, where non-linear transformations simulate realistic faults in individual modalities. Experiments on industrial datasets demonstrate that our approach achieves superior reconstruction accuracy and robustness compared to existing methods, even in the presence of corrupted or missing modalities.

Drone delivery systems are increasingly gaining importance in modern logistics due to their efficiency and potential to revolutionize the supply chain. However, among the various stages of drone operations, the landing phase stands out as the most critical, as it requires precise navigation and control to ensure the safe and successful delivery of goods. Therefore, testing simulations are required to accurately replicate the complex dynamics of real-world environments. In this paper, we propose a CycleGAN-based framework for bidirectional domain translation between simulated and real-world data, with a focus on GPS features. The proposed architecture facilitates seamless transfer learning by enabling simulation-trained controllers to operate effectively in real-world environments and vice versa. Additionally, we integrate the translated data into a downstream control framework for autonomous landing tasks, validating the practicality of the approach.

This paper presents the novel RLPID architecture for online-adaptive control, which combines classical proportional-integral-derivative (PID) control with reinforcement learning (RL). This hybrid approach enables dynamic online adjustment of PID parameters during control operation. Specifically, we propose a multi-objective reward structure that integrates established control criteria and analyze suitable configurations for different system dynamics. The RLPID controller has been implemented within the open-source middleware MLPro, where it is embedded in newly developed sub-frameworks for classical and online-adaptive control. Owing to its hybrid nature, the architecture can be used both in traditional control loops and within the Markov decision process of RL. Its effectiveness and practical applicability are demonstrated in a cascade control scenario.

Sequential addition of tasks in a learning regime leads to catastrophic forgetting of previous knowledge when learning is focused solely on new tasks. This complication can be alleviated by preserving the old data and retraining the network or introducing new networks for corresponding tasks. However, this leads to computation and storage overhead, rendering these solutions impractical in real industrial applications. We introduce a novel Gradient Monitoring (GM) approach to learning incoming tasks and simultaneously preserving old knowledge for predicting the remaining useful life (RUL) of industrial systems. A criterion matrix is formulated to evaluate the relationship between the weights of a network with their corresponding gradients generated during backpropagation. The mean of the criterion, multiplied by a learning factor, generates the threshold for creating a binary masking matrix. We propose two sub-methods of GM namely, Vanilla (VGM) and Momentum (MGM). The binary mask is utilized in VGM whereas, MGM uses the masking memory from previous training iterations along with the binary masks. We perform joint training of task-specific heads using a distillation loss and eliminate the dependency on old datasets. Using theoretical proofs, we demonstrate that GM can greatly reduce backward forgetting of old tasks. We focus our experiments on the RUL estimation of industrial systems and compare our methods with existing frameworks in continual learning (CL). For our CL experiments in RUL estimation, we select the CMAPSS Dataset and the NASA Li-ion Battery dataset, as these datasets offer subsets with varying operational features and diversity in data distribution.

We propose a novel quaternionic time series compression methodology where we divide a long time series into segments of data, extract the min, max, mean and standard deviation of these chunks as representative features and encapsulate them in a quaternion, yielding a quaternion valued time series. This time series is processed using quaternion valued neural network layers, where we aim to preserve the relation between these features through the usage of the Hamilton product. To train this quaternion neural network, we derive quaternion backpropagation employing the GHR calculus, which is required for a valid product and chain rule in quaternion space. Furthermore, we investigate the connection between the derived update rules and automatic differentiation. We apply our proposed compression method on the Tennessee Eastman Dataset, where we perform fault classification using the compressed data in two settings: a fully supervised one and in a semi supervised, contrastive learning setting. Both times, we were able to outperform real valued counterparts as well as two baseline models: one with the uncompressed time series as the input and the other with a regular downsampling using the mean. Further, we could improve the classification benchmark set by SimCLR-TS from 81.43% to 83.90%.

This article presents a novel game-theoretical (GT) framework, distributed Stackelberg strategies in state-based potential games (DS2-SbPGs), for autonomous multiobjective optimization in decentralized manufacturing systems. Existing approaches, including multiagent reinforcement learning (MARL) and native SbPG, struggle with scalability, coordination inefficiencies, and the complexity of tuning combined objective functions in real-world settings. DS2-SbPG integrates potential games and Stackelberg games, which improves the cooperative tradeoff capabilities of potential games and the multiobjective optimization handling by Stackelberg games. Notably, all training procedures are conducted in a fully distributed manner. DS2-SbPG offers a promising solution to finding optimal tradeoffs between objectives by eliminating the complexities of setting up combined objective optimization functions for individual players in self-learning domains, particularly in real-world industrial settings with diverse and numerous objectives between the subsystems. We formally prove that DS2-SbPG constitutes a dynamic potential game with guaranteed convergence. Experimental validation on a laboratory-scale testbed demonstrates the effectiveness of DS2-SbPG and its two variants: one with a single-leader–follower structure and another (Stack DS2-SbPG) for multileader–follower scenarios. Both variants significantly outperform native SbPG, achieving up to 10.61% reduction in power consumption while enhancing overall system performance, which signals the potential of DS2-SbPG in real-world applications.

In this work we propose a novel approach based on model-based Reinforcement Learning (RL) for the sim-to-real transfer of industrial robots. Specifically, we propose to employ a recently developed kinematics-informed, modular neural network serving as a learnable environment model within the world model framework. Using the kinematics-informed model, training of the world model is made more efficient resulting in faster training. Furthermore, the approach allows to train industrial robots on specific tasks solely within the simulation of the system thereby saving time and energy-consumption. Using simulations ensures safe and controlled training implementation and allows for parallelization to increase training speed. We conduct various experiments which underline the effectiveness of the proposed method. We show that training the RL algorithm solely within the simulation, results in a hundred percent task completion rate in both simulation and real world experiments.

Understanding and interpreting complex coupled systems remains one of the biggest challenges in the world. Examples of these applications range from industrial manufacturing to the temporal characteristics of real-world conditions. To address this challenge, this paper presents a novel approach to structured program code development based on graph generation. The problem is considered from the perspective of an inductive link prediction problem structured by an evolutionary algorithm. The self-adaptation of relevant knowledge takes place in a closed loop, where systematic relationships are constantly improved and extended. The required system behaviour is mapped step by step, taking into account constraints, limitations and expert knowledge. Structured graph generation is used to represent logic functions and interpret complex coupled relationships. The presented strategy enables targeted plant control through customised program code concepts, which are used to optimise processes and increase efficiency. The approach is validated through the design of interpretable programmable control logic on an industrial manufacturing process and obtain a comparable solution to the work of a trained professional. The achieved results demonstrate the next level of independent self-optimisation in learning and interpreting logical relationships in automation.

In this paper, we introduce novel gradient-based optimization methods for state-based potential games (SbPGs) within self-learning distributed production systems. SbPGs are recognised for their efficacy in enabling self-optimizing distributed multi-agent systems and offer a proven convergence guarantee, which facilitates collaborative player efforts towards global objectives. Our study strives to replace conventional ad-hoc random exploration-based learning in SbPGs with contemporary gradient-based approaches, which aim for faster convergence and smoother exploration dynamics, thereby shortening training duration while upholding the efficacy of SbPGs. Moreover, we propose three distinct variants for estimating the objective function of gradient-based learning, each developed to suit the unique characteristics of the systems under consideration. To validate our methodology, we apply it to a laboratory testbed, namely Bulk Good Laboratory Plant, which represents a smart and flexible distributed multi-agent production system. The incorporation of gradient-based learning in SbPGs reduces training times and achieves more optimal policies than its baseline.

In this paper, we propose novel quaternion activation functions where we modify either the quaternion magnitude or the phase, as an alternative to the commonly used split activation functions. We define criteria that are relevant for quaternion activation functions, and subsequently we propose our novel activation functions based on this analysis. Instead of applying a known activation function like the ReLU or Tanh on the quaternion elements separately, these activation functions consider the quaternion properties and respect the quaternion space H . In particular, all quaternion components are utilized to calculate all output components, carrying out the benefit of the Hamilton product in e.g. the quaternion convolution to the activation functions. The proposed activation functions can be incorporated in arbitrary quaternion valued neural networks trained with gradient descent techniques. We further discuss the derivatives of the proposed activation functions where we observe beneficial properties for the activation functions affecting the phase. Specifically, they prove to be sensitive on basically the whole input range, thus improved gradient flow can be expected. We provide an elaborate experimental evaluation of our proposed quaternion activation functions including comparison with the split ReLU and split Tanh on two image classification tasks using the CIFAR-10 and SVHN dataset. There, especially the quaternion activation functions affecting the phase consistently prove to provide better performance. • We define criteria for activation functions in quaternion neural networks. • We propose novel quaternion activation functions based on them. • These proposed activation functions modify either the quaternion magnitude or phase. • We investigate the gradients of these activation functions using the GHR Calculus. • We evaluate our proposed activation functions using the CIFAR-10 and SVHN dataset.

In this research, we discuss the impact of combining model-based deep learning and game theory in dynamic games to develop a sample-efficient self-learning methodology for smart production systems. We propose a novel approach, namely Model-Based Game Theory (MBGT), by incorporating model-based deep learning into a successful self-learning strategy of State-based Potential Games. Most of the advanced self-learning approaches are established based on machine learning and artificial intelligence, including reinforcement learning and game-theoretical method. However, because of the iterative behaviour during the learning processes, the learning processes are almost impossible to be conducted directly in a real-world manufacturing environment. One potential approach is to conduct the learning processes using the digital representation of the systems, e.g. the digital twin or simulation. Nonetheless, such representations in industrial settings are not always available and developing a digital representation of a complex system is immensely challenging. Hence, this problem can be solved through model-based learning, in which deep learning models are trained to predict the dynamics of the systems with a small margin of error. Then, the trained models are deployed as virtual environments for the learning processes of the GT-based control policies, instead of learning in the real world. In this study, we implement MBGT to a bulk good system and thoroughly analyse the performances between training in the real system and virtual environment, which eventually turn out to be positive effects.

Deep networks have been successfully applied to industrial applications for clean unimodal data (e.g., sensors, images, or audio). Leveraging multimodal data is a common approach to enhance performance, guided by the principle that a larger quantity of data leads to improvement. However, performance may decline considerably if corruption in the data is present (e.g., noise, blur, failure). Although researchers have explored various data augmentation methods to improve the generalization capacity, these methods are not adapted for industrial settings. The primary distinction is that current augmentation methods are designed to enhance model generalization capabilities and not realistically simulate real-world industry scenarios. In this paper, we present industry-related augmentation methods for temporal and spatial data for multimodal fusion with deep neural networks. Our methods are specifically designed to encourage modality collaboration and reinforce generalization capability. The impact of the proposed data extension strategy to train multimodal fusion models is assessed on a synthetic dataset from an industrial UR5 robot with varying degrees of imbalance. In our study, we analyze different combinations of methods and evaluate their performance. Through these experiments, we are able to identify the challenges in multimodal fusion with deep learning models in an industrial setting.

Deep networks have been successfully applied to unsupervised feature learning and supervised classification and regression for unimodal data (e.g., sensors, images, or audio). Multimodal data is often used to improve the performance of networks according to the slogan: the more, the better. Limited research is available to compensate for corrupted signals from multimodal approaches. In this work, we propose a novel regularization method for deep networks to learn features over multiple modalities designed to compensate for relative sensor weaknesses, such as sensor malfunction, inaccuracy, restricted spatial coverage, and uncertainty. We have used a special augmentation strategy for image and time series modalities to enhance the dataset of underrepresented industrial failure cases. The primary objective is to prevent these cases from negatively impacting the model's predictions. Our approach involves incorporating a fuzzy regularizer that can modify the intensity of activations depending on the quality of the signal, enabling disturbances from various modalities to be identified based on the activations. Our experiments on a simulated Universal Robots UR5 dataset demonstrate the effectiveness of our proposed regularization in increasing the model's stability, accuracy, and generalization to uncertainties and failures in the input signals. By incorporating fuzzy regularization in deep fusion models, their efficiency on complex tasks can be improved while reducing the complexity of the architectures.

The integration of advanced robotic technologies with modern software platforms has enabled significant advancements in industrial automation. This paper aims to introduce a novel workflow to integrate ABB industrial robot manipulators and the Robot Operating System (ROS), where communication is built through OPC Unified Architecture or a virtual controller with EGM+RWS. Moreover, we also provide a method to deploy the ROS environment using container technology in our workflow. The developed workflow was tested and evaluated in a real-world welding application, demonstrating its capability to automate the welding process of steel looped hooks for supermarket shelves with high accuracy and efficiency, in which the path planner for the robot is obtained from MoveIt by ROS. The introduced workflow can also be integrated with the MLPro framework, which enriches the workflow with machine learning-related algorithms, e.g. for path planning. The results of this study demonstrate the potential of integrating ABB industrial robot manipulators with ROS for industrial process automation and highlight the benefits of using ROS as a software platform for industrial robotics applications.

The digital transformation of automation places new demands on data acquisition and processing in industrial processes. Logical relationships between acquired data and cyclic process sequences must be correctly interpreted and evaluated. To solve this problem, a novel approach based on evolutionary algorithms is proposed to self optimise the system logic of complex processes. Based on the genetic results, a programme code for the system implementation is derived by decoding the solution. This is achieved by a flexible system structure with an upstream, intermediate and downstream unit. In the intermediate unit, a directed learning process interacts with a system replica and an evaluation function in a closed loop. The code generation strategy is represented by redundancy and priority, sequencing and performance derivation. The presented approach is evaluated on an industrial liquid station process subject to a multi-objective optimisation problem.

In autonomous driving, millions of frames with various scenarios for training deep object detectors is required. Labeling such a large number of frames is a costly process, therefore additional data sources support the training task. However, domain gaps from different cameras, weather, or locations typically limit the performance.We apply semi-supervised object detection, which leverages labeled source and pseudo-labeled target domain data in an iterative training paradigm. In addition, we newly include state-of-the-art adversarial style transfer into the semi-supervised training by stylizing images from source and target domains. This reduces the domain gap and improves pseudo-label quality in cross-domain semi-supervised training.In experiments and ablation studies, we show that our novel training framework can improve state-of-the-art detection performance by up to +10.1% on standard domain adaptation benchmarks.

In this paper, we present a comprehensive approach for designing and analyzing control systems with minmax constraint controllers and machine learning-based virtual sensors. By leveraging the Standard Nonlinear Operator Form (SNOF), we establish the necessary conditions for global asymptotic stability and demonstrate their applicability through an illustrative example based on a modified plant model from the literature. The proposed methodology effectively handles non-linearities and constraints, ensuring stability while providing a systematic procedure for constructing a well-formed SNOF by integrating the plant, virtual sensor, and controller. The successful application of this method in the example highlights its potential for addressing complex control problems involving min-max constraints and virtual sensors in real-world scenarios. This paper contributes to the growing body of knowledge in this area and sets the stage for future advancements, including the exploration of transforming other machine learning architectures into the SNOF and extending the stability analysis to accommodate different types of nonlinearities and constraints.

This article describes a novel concept to optimize manufacturing systems distributively through data-based learning. We propose a game-theoretic (GT) learning set-up that is incorporated with accessible control code of the programmable logic controller (PLC) to accelerate the optimal policies learning procedures, instead of learning everything from scratch. Therefore, we offer to process the accessible and available control code into a GT-based learning framework which is subsequently optimized in a fully distributed manner. To this end, we employ the recently developed framework of state-based potential games (PGs) and prove that under mild conditions PLC-informed (PLCi) learning forms a state-based PG framework. We conduct the experiment on a laboratory scale testbed in numerous production scenarios. The experiment's results highlight the major potential of using the PLCi GT-learning, which is the reduction of energy consumption of the production timescales and improvement of production efficiency while nearly halven the learning times.

This article presents a novel neural network training approach for faster convergence and better generalization abilities in deep reinforcement learning (RL). Particularly, we focus on the enhancement of training and evaluation performance in RL algorithms by systematically reducing gradient's variance and, thereby, providing a more targeted learning process. The proposed method, which we term gradient monitoring (GM), is a method to steer the learning in the weight parameters of a neural network based on the dynamic development and feedback from the training process itself. We propose different variants of the GM method that we prove to increase the underlying performance of the model. One of the proposed variants, momentum with GM (M-WGM), allows for a continuous adjustment of the quantum of backpropagated gradients in the network based on certain learning parameters. We further enhance the method with the adaptive M-WGM (AM-WGM) method, which allows for automatic adjustment between focused learning of certain weights versus more dispersed learning depending on the feedback from the rewards collected. As a by-product, it also allows for automatic derivation of the required deep network sizes during training as the method automatically freezes trained weights. The method is applied to two discrete (real-world multirobot coordination problems and Atari games) and one continuous control task (MuJoCo) using advantage actor-critic (A2C) and proximal policy optimization (PPO), respectively. The results obtained particularly underline the applicability and performance improvements of the methods in terms of generalization capability.

In Machine Learning, large models need to have access to a huge amount of training data. This requirement applies to many applications in an industrial environment. Furthermore, in specific processes it is not easy to obtain such a large amount of data for different reasons ranging from privacy, security, and even process affectations. Therefore, this work proposes the creation of synthetic time series datasets to simulate processes out of a given subset of functional relationships. Moreover, using transfer learning to improve the performance of four autoencoder architectures in terms of unsupervised time series reconstruction, requiring fewer target data for training. We outline multiple concepts of data generation and use statistical analysis to evaluate the dataset performance and complexity. Further, the data is used to train unsupervised models and enables them to improve their reconstruction performance over 52 sensors and multiple fault cases. By reducing the amount of available train data, we still gain sufficient results through the pre-training. Overall, we see significant performance and interpretability improvements on a new time series analysis approach named Bag-of-Functions compared to convolutional and linear autoencoders.

This paper presents a novel approach to train a Reinforcement Learning (RL) agent faster for transportation of parcels in a Peristaltic Sortation Machine (PSM) using curriculum learning (CL). The PSM was developed as a means to transport parcels using an actuator and a flexible film where a RL agent is trained to control the actuator. In a previous paper, training of the actuator was done on a Discrete Element Method (DEM) simulation environment of the PSM developed using an open-source DEM library called LIGGGHTS, which reduced the training time of the transportation task compared to the real machine. But it still took days to train the agent. The objective of this paper is to reduce the training time to hours. To overcome this problem, we developed a faster but lower fidelity python simulation environment (PSE) capable of simulating the transportation task of PSM. And we used it with a curriculum learning approach to accelerate training the agent in the transportation process. The RL agent is trained in two steps in the PSE: 1. with a fixed set of goal positions, 2. with randomized goal positions. Additionally, we also use Gradient Monitoring (GM), a gradient regularization method, which provides additional trust region constraints in the policy updates of the RL agent when switching between tasks. The agent so trained is then deployed and tested in the DEM environment where the agent has not been trained before. The results obtained show that the RL agent trained using CL and PSE successfully completes the tasks in the DEM environment without any loss in performance, while using only a fraction of the training time (1.87%) per episode. This will allow for faster prototyping of algorithms to be tested on the PSM in future.

In this paper, we discuss the impact of communication and memory-based learners on distributed self-optimization of smart and flexible manufacturing units. Specifically, we employ the recently proposed framework of state-based potential games, which has proven to be successful in allowing distributed optimization in multi-agent systems. We first augment the framework with additional communication capabilities for the individual players and analyze the efficacy of state and action communications within the different players. Second, we incorporate memory states within the learning dynamics of the players and analyze their impact on the learning performance. The proposed method is inspired by the promising results of memory-based reinforcement learning. However, previous studies have rarely dealt with distributed manufacturing control. We believe that it will be important to explore the potential use of the communication and memory-based approaches in manufacturing control with multi-agent settings. Hence, the proposed method is applied to a bulk good laboratory plant providing a thorough experimental analysis of the effect of the various improvements with very encouraging results.

This article presents a novel approach for distributed optimization of production units based on potential game (PG) theory and machine learning. The core of our approach is split into two parts: the first part concentrates on the conceptual treatment of modular installed production units in terms of a PG scenario. The second part focuses on the development and incorporation of suitable learning algorithms to finally form an intelligent autonomous system. In this context, we model the production environment as a state-based PG where each actuator of each module has the role of an agent in the game aiming to maximize its utility value by learning the optimal process behavior. The benefit of the additional state information is visible in the performance of the algorithm making the environment dynamic and serving as a connector between the players. We propose a novel learning algorithm based on a global interpolation method that is applied to a laboratory scale modular bulk good system. The thorough analysis of the encouraging results yields to highly interesting insights into the learning dynamics and the process itself. The benefits of our distributed optimization approach are the plug-and-play functionality, the online capability, fast adaption to changing production requirements, and the possibility of an IEC 61131 conforming to PLC implementation.

Analyzing health records to avert future complications and provide the right treatment plays an important role in medical diagnostics. This paper introduces a method to use real-life ICD-coded Electronic Medical Records (EMR), clinical data collected between 2018-2019 in Africa, to create a prediction model. The prediction model uses recent advances in machine learning, specifically Convolutional Neural Networks (CNN), to provide an alternative and a powerful way for disease risk prediction. An algorithm is first designed to process the real-life EMR data into multi-step time-series forecasting data. A CNN model is then developed, to predict the individual’s future disease class (ICD-10 CM chapters) risk based on their demographic and medical history. The experimental results show that the model predicts future disease risk, for an individual, with an accuracy of 80.73%. We conclude that such a model can prove immensely useful in cost savings for individuals and hospitals by performing preemptive corrective action, provide additional guidelines for hospitals’ capacity planning, when the model is applied for demography, and, most importantly, peace of mind on personal health.

Remaining useful lifetime (RUL) estimation of engineering assets is of great importance in an efficient operation of a production system. This paper deals with the RUL estimation with deep convolutional neural networks (DCNN) of bearings being operated in different load conditions by measuring the horizontal and vertical vibration. We present the use of Sobolev training to enhance the prediction capabilities of the DCNN since deeper architectures generally perform better than their shallower counterparts. We test the proposed approach on the benchmark IEEE PHM 2012 data challenge datasets for RUL prediction and compare the results with standard DCNN approaches. The results show that the proposed training methodology consistently outperform the standard training approach for RUL estimation even though the model size and complexity was much lower.

This paper presents a novel approach to incorporate detailed Discrete Element Method (DEM) simulation models into a data-based control system applied to the control of a peristaltic singulation and sortation machine (PSM). The bionic principle of peristaltic is often used in nature for locomotion or transporting of goods. For sortation and singulation of parcels, peristaltic movements provide the advantage of faster operation due to the potential parallelization of the singulation and sortation processes and a far more gentle parcel transport. Beside the mechanical design, a major challenge of the PSM design lies in the development of suitable control algorithms. Due to the difficulties to model the complex behaviour of the independent parts physically, the design process appears to be hardly possible and thus a data-based control design based on reinforcement learning (RL) is proposed. Further, a co-simulation which incorporates a detailed DEM simulation is developed. Particularly, by suitably combining the state-of-the-art actor-critic reinforcement learning (ACRL) and a distributed approach using multiple parallel environments, manageable simulation and training times are ensured. The obtained results show the applicability of a DEM model in a co-simulation framework solving the transportation problem of parcels and also the very good performance of the developed RL-based control approach.

The constant and rapid rise in the field of Industrial Internet of Things has enabled the manufacturing and process industries to have access to large amounts of process data. This process data can be effectively analyzed to identify the faults in the system thereby facilitating in avoiding critical process breakdowns. Deep neural networks with their inherent ability to model complex non-linear representations, have been proven to fit well for contemporary fault detection. This work proposes a Time Series based approach to fault detection in the benchmark Tennessee Eastman process making use of the temporal dependencies within the process data. Since standard Feed Forward Neural Networks (FFNN) are not capable of learning these temporal dependencies, a novel approach using Convolutional Neural Networks (CNN) with its architectural and algorithmic variants is proposed. The experimental results show comparatively superior performance of the proposed CNN based models with the standard FFNN for fault detection. Also the different hyperpaprameters which effect the time series classification task are highlighted.

Early fault detection and diagnosis lead to enhanced production performance since the companies can avoid unplanned machines downtimes and final product quality issues. Expert systems are widely popular in the industry for fault detection and diagnosis applications. This is due to its easy implementation, and its versatility to transform expert knowledge into rules. This paper proposes an extended FMEA (eFMEA) as a systematic knowledge extraction tool to detail the failure causes of equipment in an industrial process. It outlines a methodology for the automatic generation of expert knowledge rules through the digitalisation of the eFMEA. The last contribution is a rule-based expert system for multi-fault detection and diagnosis based on the digitalised eFMEA. This expert system uses the communication protocol OPC-UA to feed the current data from the equipment, to detect faults. It can detect multi-faults simultaneously. The fault detection and diagnosis are visualised through a web application.

Anomaly detection methods are used for fast and reliable detection of abnormal events in industrial processes. The early detection of anomalies can avoid critical process breakdowns and hence can increase the overall productivity of the system. The availability of labelled datasets for all the possible faulty scenarios is generally not possible, as most of the industrial systems operate in a non-faulty condition. Deep learning architectures that can be trained in an unsupervised setting such as deep autoencoders, denoising autoencoder and variational autoencoder provide an appropriate solution to this problem of unlabelled data for industrial anomaly detection. We investigate and compare the applicability of these architectures on the benchmark Tennessee Eastman fault detection study. The deep architectures are trained to model only the normal operating condition with its threshold set by kernel density estimation. A detailed comparison from the experimental results shows superior anomaly detection capabilities of the variational autoencoder as compared to the other methods.

We present a novel approach for distributed optimization of highly flexible, modular production units enabling plug-and-play production with online optimization capabilities to adapt fast to changing production requirements. The approach is fully distributed in the sense that each production module to be optimized is equipped with its own optimization agent which local objective is the optimization of its own production objectives. To assure the necessary coordination between the agents, the resulting distributed optimization problem is designed using concepts of game theory. To this end, we model the production environment in terms of a potential game where each module is modeled as a player of the game. By assigning suitable utility functions to the players coordination of the agents behavior is achieved to find an optimal collective behavior. We apply the approach to a laboratory scale distributed bulk good production testbed with very encouraging results. In addition, due to the computational simplicity of the approach, an implementation in IEC61131 compatible code is possible allowing a direct implementation of the approach in existing production units.

This paper introduces a novel approach for the control of flexible manufacturing units by means of cooperatively interacting industrial robots. For fast adoption to the actual production requirements, we embed a learning module into the manufacturing cell. This learning module allows the robots to learn to solve the given task with respect to a given optimization objective. Simultaneously, the robots learn to efficiently cooperate and find an optimal collective behavior while solving the task. To this end, we develop two different control algorithms based on reinforcement learning. The first approach is based on a centralized agent which coordinates the learning behavior of the whole manufacturing cell. In the second approach, a learning agent is assigned to each robot allowing for more flexibility and reducing the state-action space of the reinforcement learning problem at hand. The approaches are applied to a laboratory testbed using two cooperating industrial robots which should learn to optimize the throughput of the manufacturing cell. A comparison of both approaches shows the improved performance of the decentralized learning agents compared to the centralized one both in goal achievement and learning speed.

This paper introduces a novel approach for self-learning in highly flexible, modular manufacturing systems enabling fast reconfiguration and online adaptation to changing production requirements. The approach is based on a distributed optimization scheme such that production modules are equipped with their own optimization agent with its local objectives to be optimized. The communication and coordination of the agent is limited to the basically required amount. The approach is based on the recently developed deep deterministic policy gradient (DDPG) approach, a high performing algorithm from the family of actor-critic reinforcement learning algorithms. As DDPG is based on single agent learning, we develop a fully distributed multi-agent learning setting with different levels of information about the neighbors. We apply the approach to a laboratory scale distributed bulk good production testbed with very encouraging results. Particularly, we found very reasonable control strategies by learning the agents from scratch.

Classifiers based on machine learning are popular in literature, in order to support predictive maintenance of machinery. Depending on the process data, one classifier can assess target classes better than others. It often happens that the classifiers complement each other. A fusion strategy is needed in order to exploit the strength of each classifier. This paper presents fault detection assessment architectures based on information fusion and classification methods. It proposes the use of information fusion methods and different architectures, in order to improve the overall result of the fault detection assessment. Dempster-Shafer and Yager rules of combination are used to fuse the classification method predictions. The rules of combination improve the results by complementing the classifiers performance. A comparison between centralized and decentralized architectures is presented. The results show that the information fusion using decentralized architectures improves the overall performance of the fault detection assessment.

In the framework of the digitalization of the industry, there is an increasing trend to use machine learning techniques to assess condition monitoring, fault detection, and process optimization. Traditional approaches use a local Information Technology (IT) framework centralized in a server in order to provide these services. Cost of equipment and IT manpower are associated with the implementation of a condition monitoring based on machine learning. Nowadays, cloud computing can replace local IT frameworks with a remote service, which can be paid according to the customer needs. This paper proposes a cloud-based architecture for condition monitoring based on machine learning, which the end-user can assess through a web application. The condition monitoring is implemented using a fusion of classification methods. The fusion is implemented using Dempster-Shafer Evidence Theory (DSET). The results show that the use of DSET improves the overall result.

Machine faults cause high financial losses for industrial companies. Machine faults stop the process flow, and are associated with idle time and additional expenses due to unexpected corrective maintenance. Depending on the fault, the idle time can span from short time actions to long inactivity periods for the machine. This idle time can risk a customer shipment, subsequently putting at risk the customer trust. A complex interaction between process variables cannot be easily traced using conventional condition monitoring approaches. There is an increasing trend to use machine learning methods to assess condition based maintenance. This paper proposes a mobile App for the fault detection assessment of industrial machinery. A proposal for the design and implementation of the mobile is presented, as well as the methodology followed for the fault detection assessment. The fault detection assessment is adressed using machine learning techniques, specifically a majority voting classifier. This majority voting classifier combines the predictions of three classification methods. The results show that the majority voting approach improves the performance of the fault detection assessment in comparison with the invidual classification methods.

This paper presents a novel virtual commissioning approach based on the Discrete Element Method. To save costs and reduce time at the commissioning phase, this technique offers a suitable tool to model complex behaviour of bulk material handling. This proposed approach of virtual commissioning is used for validating existing software code and solve an optimisation problem, which yields to an optimal throughput of a bulk good handling process by only using the simulated hardware representation. Due to that, this work shows the complete framework to apply the Discrete Element Method as a hardware-in-the-loop system for virtual commissioning. The steps from the design of mechanical parts, the model of the bulk good material as well as the communication structure is proposed. The whole approach is compared with a conventional hardware-in-the-loop system regarding the required engineering efforts in different disciplines. It is pointed out that the novel approach of a virtual commissioning based on Discrete Element Method is a highly useful tool beyond the functionalities of common virtual commissioning approaches.

Industrial companies are constantly looking for methods to maintain a certain process efficiency and steady performance of their machines. Assistance systems support users in different applications such as driving, industrial expert systems, and physician assistance to achieve these objectives. This paper proposes an assistance system, which can support an operator by means of using the accumulated knowledge of process experts, and thus providing recommendations for improving the process performance. An architecture based on information fusion and rule-based programming is presented. The information fusion considers the Dempster-Shafer rule of combination, in order to fuse the process knowledge of different experts into a consolidated expertise. A rule-based approach serves as a way to combine the process knowledge with the machine data. The technical implementation considers an industrial communication protocol OPC-UA, which enables a data interface between the assistance system and the automation platform. This paper presents an application case to illustrate how the assistance system interacts with the process experts, as well as the assessment to the operator.

Process monitoring and fault diagnosis methods are used to detect abnormal events in industrial processes. Process breakdowns hinder the overall productivity of the system which makes the early detection of faults very critical. Due to the highly non-linear nature of modern industrial processes, deep neural networks with several layers of non-linear complex representations fit aptly for contemporary fault diagnosis. Although deep neural networks have found wide array of application areas such as image recognition and speech recognition, their effectiveness in fault detection has not been tested substantially. In this study, a comparison between two deep neural network architectures, namely Deep Stacking Networks and Sparse Stacked Autoencoders for fault detection from process data is presented. The Tennessee Eastman benchmark process is considered to test the effectiveness of these deep architectures. A detailed comparison between the two architectures is illustrated with different hyperparameters. The experiment results show that the Sparse Stacked Autoencoders model has superior average fault detection capability and is also more stable as it has less variation in fault detection rate.

This paper presents a novel approach to operate industrial robots as used for manufacturing lines within a cooperative robot station. The proposed framework consists of the application of especially to the cooperative robot handling problem adjusted Reinforcement Learning (RL) algorithms. Such RL-algorithms deal with sequential decision making processes in a trial-and-error learning interaction with the environment, to finally gain an optimal team-working behavior among the robots. In particular application results to a real team-working robot station underline the effectiveness of the novel RL approach.

This paper presents a novel approach for self-reconfiguration and plug-and-play control of modular production units. The approach is based on a service-oriented architecture which allows for a fully decentralized control and communication structure. Each system module is equipped with its own control system and communicates with other units by means of predefined service requests. Hence, the control of the overall system is solely executed by the modules own control and by the exchange of service requests of each station. The approach is implemented on a laboratory bulk good system typically used in the pharmaceutical industry with modular system structure. Results obtained from experiments underline the applicability of the approach.

This paper presents a novel approach to the optimization of energy consumption in large scale industrial bulk good processes. The approach is based on a model-free self-learning algorithm solely based on available process data using ideas from the well known reinforcement learning framework. To this end energy consumers of the plant are integrated in the optimization framework such that each consumer learns its own optimal energy profile for a given production task. The approach is implemented on a laboratory size testbed where the task is the supply of bulk good to a subsequent dosing section. The capability of the approach is underlined by the results obtained at the testbed.

This paper presents a new approach to the stability analysis of recurrent fuzzy systems (RFSs). RFSs are rule-based dynamic fuzzy systems that are usually obtained from heuristic or data-driven modeling. In the presented approach, the stability of both continuous-time and discrete-time RFS can be analyzed in a common framework. It is based on the representation of an RFS as a hybrid polynomial system. Due to the polynomial structure, the recently developed method of sum-of-squares (SOS) decomposition along with semidefinite programming can be employed to derive sufficient conditions for the stability of equilibrium points. We consider stability analysis for known equilibrium points as well as unknown equilibrium points. The latter results in a two-step procedure. In the first step, a polynomial is constructed. In the second verification step, this polynomial is proven to be a Lyapunov function for the RFS. The applicability of the approach is shown by two systems formulated as a rule-based RFS.

This paper proposes a new approach to system modeling using continuous-time recurrent fuzzy systems (CTRFS). The approach is based on the representation of CTRFS as hybrid systems. With this representation, various forms of a priori knowledge about the system to be modeled can be incorporated. This allows a reasonable reduction of optimization parameters and hence, avoids overfitting. Furthermore, the presented approach offers a deeper insight into the system structure on the basis of measurement data solely. This is illustrated by a selection algorithm for relevant input and state variables. The applicability of the approach is shown by modeling a chemical process.

In this paper, we present a new approach for the stability analysis of continuous-time recurrent fuzzy systems (CTRFS). The approach is based on the representation of a CTRFS as a switched polynomial system, for which a Lyapunov function is constructed in a two step procedure. Both steps are based on semidefinite programming. The applicability is shown by an ecological system formulated as a rule based structure.