Accurate and robust detection of anomalies in time series is fundamental to many established methods of intelligent automation. For example, accurate anomaly detection is a prerequisite for successful predictive maintenance or security monitoring. In the recent past, neural networks built in autoencoder structures have become the state of the art in anomaly detection. A new network architecture in deep learning — the Kolmogorov-Arnold networks — coupled with wavelet-based activation functions (wavKAN) currently promises better performance in time series analysis than classical multilayer perceptrons (MLP). This paper evaluates wavKAN-based autoencoders on the UCR Time Series Anomaly Archive, a repository of 250 datasets including industrially relevant applications. The results of the evaluation show that wavKAN-based autoencoders detect anomalies more accurately than MLP-based autoencoders, while relying on a smaller number of parameters. WavKANs thus offer a novel approach to anomaly detection that reduces hardware requirements in industrial applications, making them particularly suitable for edge devices.

Many small and medium-sized enterprises lack large datasets and AI expertise, limiting their ability to apply traditional AI methods. However, they often possess valuable yet underutilised experimental data. This paper introduces an interpretable AI workflow tailored for such scarce data environments. It guides users through experimental design, data labelling, Decision Trees, and Active Learning to optimise processes efficiently. A bread roll baking use case illustrates the workflow’s practical value and transferability to other industrial settings.

The production of plastic fibers is a complex process, where the quality of the produced fibers is crucial for their performance in applications like fire-resistant concrete. Their quality is determined through visual inspections of dispersion tests, which are chemically treated samples of the produced fibers. However, this process is prone to more conservative ratings by lab personnel, as to avoid missing existing defects. This leads to unnecessary blade replacements and increased costs. To support the lab personnel in their rating process, a machine learning based pipeline is proposed. It consists of a classification model and explainable artificial intelligence (XAI) methods to visualize the presence of agglomerations, which are the sole determinant of the dispersion test’s rating. The goals are to provide lab personnel with a second opinion on the quality of the dispersion tests, whilst having only scarce labeled data available. Additionally, the suitability of different XAI methods and their pre- and post-processing steps is evaluated. The proposed pipeline is able to classify the dispersion tests with an accuracy of 85.92 %. The results highlight Layer-wise Relevance Propagation and HiResCAM as the most suitable XAI methods for this use case.

The integration of deep learning and information fusion into Cyber-Physical Production Systems offers significant opportunities to increase manufacturing efficiency, enable real-time root cause analysis, improve process monitoring, and support quality prediction. This work-in-progress publication explores a novel approach that combines deep learning for image-based quality inspection with multi-source information fusion to optimize complex manufacturing processes, with a particular focus on injection molding. Traditionally, automated visual inspection systems and industrial condition monitoring are not self-adaptive and are operated in data silos, requiring continuous intervention and parameterization by industrial engineers - often based on expert intuition rather than empirical data. This reliance on expert knowledge limits scalability and responsiveness to process variations. By merging industrial computer vision techniques with heterogeneous sensors and information data, a robust hybrid information fusion framework is proposed. This approach enables Cyber-Physical Production Systems to correlate visual defect classification with underlying process parameters, facilitating autonomous learning from increasing amounts of production data. As a result, Cyber-Physical Production Systems can dynamically adapt without manual input, leading to improved process stability and product quality. Initial proposals in this publication highlight the potential of this method to deliver practical, AI-driven improvements for industrial environments, underscoring the transformative value of intelligent and data-centric manufacturing systems.

Time series forecasting allows businesses to gain insights into market trends and plan production, storage, and sales in advance. There is a shift towards applying machine learning (ML) models for forecasting on hierarchical sales datasets. However, large historical training data is not always available, particularly in markets where sales are not frequent but rather occur weekly or monthly. In such cases, current ML approaches often require years of training data. This study focuses on small datasets and evaluates the potential of ensemble methods to improve prediction accuracy. Forecasting models are assessed on three hierarchical sales datasets from two companies. Sixteen traditional statistical and ML models were evaluated using customized test periods for each dataset. The results show that the best-performing model depends on the dataset's characteristics. A seasonal naive model showed effective seasonality detection, while a baseline model robustly outperformed complex models in one of the datasets. Among traditional statistical models, theta and exponential smoothing produced similar results, and a statistical ensemble outperformed in larger prediction horizons. ML models showed mixed results, with ML ensembles showing significant advantages on the largest dataset. The study demonstrates the importance of model choice relative to dataset specifics, with traditional statistical approaches often outperforming their machine-learning counterparts on small datasets.

Vaseline, also referred to as petrolatum, is a colloidal dispersion of liquid-crystalline structures of hydrocarbons derived from petroleum. It has long been recognized for its versatile applications in the pharmaceutical industry, with its use in the formulation of various topical medications, wound care products, and drug delivery systems. For pharmaceutical use, petrolatum has to meet the quality standards described in its Pharmacopoeia monograph. The comprised test ranges allow for a broad range of Vaseline qualities on the market, while the tests themselves only poorly discriminate between grades. The only differentiating properties are related to the melting behavior, which is tested via drop point analysis, and the consistency, addressed in the functionality-related characteristics section. In this study, we propose the hypothesis that Near-infrared spectroscopy (NIRS) could be a comparably simple method to evaluate the crystalline behavior of Vaseline qualities. We expect such information to provide additional details for Vaseline quality discrimination. This discrimination would allow the most suitable petroleum jelly to be selected for an existing formulation when the previous one needs to be replaced; for example, due to a manufacturer change. We demonstrate that NIRS in transmission and reflectance mode obtained by traditional continuous spectra acquisition and fragmented NIR spectra acquisition through multi-optical, multi-modal excitation, respectively, can both serve as a basis for detecting Vaseline quality differences, which we have further proven by thermal analysis and tests with semisolid formulations. Additionally, we demonstrate that a lower-cost multi-optical spectrometer in reflectance mode can detect Vaseline quality differences in rotated samples.

The increasing size and complexity of datasets, the need for constant adaptation to current conditions, and the potential benefits of machine learning (ML) techniques, such as flexibility and the ability to incorporate additional features, have led to the increasing use of ML techniques in forecasting as an alternative to traditional statistical methods. However, the results are often not transferable to smaller datasets. This paper analyses a real inventory management dataset and compares statistical and ML methods to determine which techniques consistently produce accurate results, even for smaller datasets. The results show that the choice of aggregation level affects the performance of statistical and ML methods, with the LightGBM model showing consistent performance across different scenarios and aggregation levels, and simpler methods effectively modelling intermittent or lumpy time series.

In the contribution at hand, multilayer feedforward neural networks based on multi-valued neurons (MLMVN) are applied on a classification problem in the context of cyber-physical systems. MLMVN are a specific type of complex valued-neural networks. The aim is to apply MLMVN on a benchmark dataset and to classify individual states of a motor (one non-fault state and 10 different fault states). For the multi-class classification problem, an evaluation of selected real-valued and complex-valued feedforward neural networks is considered. One finding is that in terms of accuracy, shallow MLMVN significantly outperform similarly constructed real-valued feedforward neural networks on the benchmark dataset. Thus, the high efficiency of such networks could be an advantage when processing data locally in order to improve robustness, performance, and reduce energy consumption on the system in use.

Detecting redundancies between information sources is essential for applications both in machine learning and information fusion. State-of-the-art redundancy metrics, such as correlation coefficients or mutual information, are based on probabilistic concepts. In technical multi-source systems information is often uncertain but also incomplete. Thus information is often provided with uncertainty distributions (probabilistic or possibilistic). In this paper a redundancy metric is proposed which is embedded in the framework of possibility theory applicable incomplete and uncertain information. The metric considers the consistency between sources, the specificity of pieces of information, and the range of observed information over the frame of discernment. The redundancy metric is designed to be cautious since incorrect identification of redundancies affects both machine learning and information fusion applications negatively. A machine learner may be deprived of information, whereas an information fusion system, relying on false assumptions, may, e.g. , incorrectly assess sources as unreliable. The proposed redundancy metric is qualitatively evaluated on information sources of three technical datasets.

In the filling and packaging industry, the trend is towards self-diagnosis, optimization, and quality monitoring of processes. The aim is to increase production volumes and the quality. These concepts require continuous monitoring and anomaly detection of the filling process. In addition, a root cause analysis of the failure is required because not every failure can be simulated or measured previously. Standard anomaly detection methods have no integrated root cause analysis. In this paper a fusion system is utilises for the detection of different unknown anomalies and also the failure source of them. The performance of this method is benchmarked with a real-word filling process.

On-site aircraft repairs are gaining in importance due to the susceptibility of carbon fibre reinforced polymers to damage. Repairs themselves are required to be inspected for quality, preferably cost- and time-efficiently. This paper presents an approach for the inspection of repaired composites based on guided Lamb waves. The focus is on cost-effective signal excitation and effective signal processing. Lamb waves are excited with piezoelectric transducers at the resonance frequency of the material under test. Measured signals are processed with a complex wavelet transform to improve damage detection. The proposed approach is evaluated on two test specimens, one of which has a defect in the adhesive bond.

The aim of sensor orchestration is to design and organise multi-sensor systems both to reduce manual design efforts and to facilitate complex sensor systems. A sensor orchestration is required to adapt to non-stationary environments, even if it is applied in streaming data scenarios where labelled data are scarce or not available. Without labels in dynamic environments, it is challenging to determine not only the accuracy of a classifier but also its reliability. This contribution proposes monitoring algorithms intended to support sensor orchestration in classification tasks in non-stationary environments. Proposed measures regard the relevance of features, the separability of classes, and the classifier's reliability. The proposed monitoring algorithms are evaluated regarding their applicability in the scope of a publicly available and synthetically created collection of datasets. It is shown that the approach (i) is able to distinguish relevant from irrelevant features, (ii) measures class separability as class representations drift through feature space, and (iii) marks a classifier as unreliable if errors in the drift-adaptation occur.

We present an approach for feature extraction in the context of condition monitoring of a bottling process. A special focus lies on the characterisation and evaluation of liquid textures. The approach will feed into a sensor and information fusion system to monitor a bottling process. Requirements like real-time capabilities, data reduction and resource limitations necessitate a fusion approach which capture physical effects of different sensors, extract appropriate features and combine them into one state for the complete filling process. Special attention is paid to the feature extraction of the visual sensor signals to monitor the filling level, the amount of foam and the degree of turbulence in the liquid.

We present a method for the fast and robust linear classification of badly conditioned data. In our considerations, badly conditioned data are such data which are numerically difficult to handle. Due to, e.g. a large number of features or a large number of objects representing classes as well as noise, outliers or incompleteness, the common software computation of the discriminating linear combination of features between classes fails or is extremely time consuming. The theoretical foundations of our approach are based on the single feature ranking, which allows fast calculation of the approximative initial classification boundary. For the increasing of classification accuracy of this boundary, the refinement is performed in the lower dimensional space. Our approach is tested on several datasets from UCI Reposi-tiory. Experimental results indicate high classification accuracy of the approach. For the modern real industrial applications such a method is especially suitable in the Cyber-Physical-System environments and provides a part of the workflow for the automated classifier design.

Multisensor systems are susceptible to sensor ageing effects as well as to environmental changes. Due to these effects, the distribution of sensor measurements may change over time, which is referred to as sensor drift. A multisensor system which adapts to drift by self-monitoring is more durable, requires less manual maintenance, and provides information of higher quality. This contribution proposes an approach for detecting and adapting to sensor drift. The proposed detection algorithm determines the reliability of a sensor based on fuzzy pattern classifiers and a consistency measure. By this means, the inherent redundancy in multisensor systems is exploited to detect drift. Detected drift leads then to a retraining of the classifier on batched data guided by information fusion. The retraining incorporates the estimated magnitude of the drift. The proposed algorithms are evaluated in comparison with state-of-the-art methods in the scope of a publicly available dataset. It is shown that the drift detection algorithm yields results similar to the benchmark algorithm but is less computationally complex. Relearning with the drift-adapted approach results in more robust classifiers with regard to potential future drift.

The current trend towards mass customisation requires adaptive, modular, and flexible production systems. The installation, configuration, and monitoring of such systems are becoming increasingly time-consuming, expensive, and complex tasks. The related challenges are met by self-organisation and information fusion techniques. Distributed self-organising systems are robust, scalable, and inherently modular, whereas information fusion techniques reduce the complexity of information from distributed sources. The combination of both, an automated design of information fusion systems taking advantage of self-organising methods, is an open and active research field. This contribution proposes an approach for agent-based intelligent sensor nodes which cooperate towards designing an information fusion system relying on semantic self-descriptions. The focus is on communication and making collective decisions including task allocations and election processes. The performance of the proposed approach is evaluated in comparison to a centralised state-of-the-art design system. It is shown that the proposed approach scales similarly, but due to its distributed nature, it is more robust to device failures.

Currently, new research questions arise because of the paradigms of Industry 4.0, which aims to bring together mechatronic systems and information technologies. Its general idea is to create an Internet of Things consisting of communicating machines, which implement concepts for self-configuration, -diagnosis, and -optimisation. The realisation of these functionalities is in focus of current research and gains in importance not only in the industrial sector. The overall goal is to equip technical systems with intelligence to enable for autonomous behaviour. Therefore, tasks like information processing, extensive networking, or system monitoring using sensor and information fusion systems have to be reconsidered. This contribution focuses on the design and maintenance of sensor and information fusion systems and presents a preliminary evaluation of a design concept for such applications. The concept is developed to automatically configure sensor and information fusion systems, which is a time-consuming and complex task when carried out manually. It reduces the perceived complexity of the application and supports the designer during design and maintenance of the sensor and information fusion system.

The way how we interact with banknotes is changing. This raises questions on how we interact with electronic payment systems. The general idea is to design low-cost electronics for cash handling systems. We establish a prototypical demonstrator which allows a consistent image capture quality and is able to handle complex algorithms for banknote authentication on cost-effective hardware. Therefore, tasks regarding reducing the evaluation time, without diminishing the reliability of the algorithms have to be considered. In this contribution we focus on the re-design of an authentication module for detection of commercial offset printing. This module analyses images in view to periodic printing patterns by means of the Discrete Fourier Transform (DFT). We propose to implement two concepts: an adaptive software architecture for DFT and parallel image processing. The re-design reduces evaluation time, without compromising the reliability of the authentication algorithm.

Cash machines or automated teller machines (ATMs) are one of the typical ways to get cash around the world. Such machines are under a variety of criminal attacks. Most of the manipulations are performed through skimming. In 2014, such attacks led to a damage of approx. 280 million Euro within the EU. In this paper, we propose an approach to detect anomalies and attacks on ATMs via motif discovery. Motifs are frequently unknown occurring sequences or events in a time series signal. State of the ATM is captured by innovative piezoelectric sensor networks to analyse the occurring vibrations. The captured signals are inspected by the Complex Quad-Tree Wavelet Packet transform which provides broad frequency analysis of a signal in various scales. Next, features are extracted from the selected scale based on the information content, to detect motifs. Detected motifs provide the prototype patterns for anomaly detection or classification tasks.

Cyber-Physical Production Systems (CPPSs) are in the focus of research, industry and politics: By applying new IT and new computer science solutions, production systems will become more adaptable, more resource ef- ficient and more user friendly. The analysis and diagnosis of such systems is a major part of this trend: Plants should detect automatically wear, faults and suboptimal configurations. This paper reflects the current state-of- the-art in diagnosis against the requirements of CPPSs, identifies three main gaps and gives application scenarios to outline first ideas for potential solutions to close these gaps.

Due to the material changes of components from metal to plastic or composite materials, the structural health monitoring finds more and more interest in the industrial fields. The reason is that these materials are more vulnerable to damage or impacts which cannot be optically detected. In this contribution we present a method to analyze the structure of plastic components with piezo-electrical sensors and actuators. The components are stimulated by actuators, and sensors capture the injected vibrations. These signals are decomposed into Intrinsic Mode Functions to compute statistical features. A Fuzzy-Pattern-Classifier is applied to detect structural modifications at the components under test.

In modern industrial applications driven by Cyber-physical systems (CPS) it is a challenging task to model and optimize processes such as machine analysis and diagnosis. Since the CPS have to act autonomously, a procedure for automated decision making has to be designed. In our work we concentrate on the design of a decision procedure by a fuzzy classifier approach. For our application on decision making in an industrial environment, a fuzzy approach was picked as convenient classification technique regarding balance between accuracy and computational time. We present a supervised learning method called FUZZY-ComRef which combines fuzzy classification and our combinatorial refinement method, called ComRef [1]. Due to the fact that fuzzy classification might behave inaccurately for some datasets, the aim of our approach is to improve the results provided by the (stand-alone) fuzzy classification. We show the performance of FUZZY-ComRef evaluated on the samples from the UCI Repository and on our real-world dataset Motor Drive Diagnosis. In addition, we discuss the quadratic computational time problem arising from the combinatorial nature of ComRef. Furthermore, we show based on real-time evaluations that within parallelisation the proposed FUZZY-ComRef is suitable to many applications in CPS.

We present a new approach for linear classification optimisation based on Combinatorial Refinement (ComRef) of feature weighting for cognitive signal processing in resource-limited hardware and software like in Cyber-physical systems. Despite simple construction, the approach is able to connect advantages of dimensionality reduction methods and such like combining multiple classifiers resp. Bag-of-classifiers-approaches and leads to a good generalisation ability even by use of small feature sets. Regarding generalisation ability, we benchmark the performance of ComRef on several datasets from the UCI repository. Furthermore, for an industrial dataset Motor Drive Diagnosis we show the advantage of ComRef which uses Support-Vector-Machines (SVM). In this application scenario, a trustful classifier is essential, since a small number of mis-classifications could lead to motor damages.

Many modern industrial applications, e.g. those incorporating hundreds or thousands of electrical sensors and actuators, must be categorised into Big Data environments, in which it is essential to design suitable information processing models. Central data processing in such environments is impossible and must be carried out in a distributed way on resource-limited cyber-physical systems. One of the challenging tasks for machine learning is thus the design of a classifier which is simple, accurate and has an acceptable realisation time. We present ComRef-2D-ConvHull method for linear classification optimisation in lower-dimensional feature space, which is based on ComRef from [1]. Compared to original ComRef, we consider only classification optimisation in 2-dimensional feature spaces in ComRef-2D-ConvHull. Due to the decreased time complexity for calculations in 2-dimensional feature space, we expect many industrial Big Data enviroments to profit from our method. Tests regarding the generalisation ability of ComRef-2D-ConvHull on several reference data sets and on a real-world industrial dataset show promising results.

In this contribution we show enhancements of the safety of hazardous material stores by the usage of a condition monitoring system. Hazardous material stores function as a store for dangerous chemicals. We use fire simulations to simulate a fire and use the results of this simulation in our condition monitoring system in order to show the attainable gains. The used condition monitoring system utilises multiple sensors which are distributed inside and outside of the hazardous material store. The values of the sensors are combined over multiple levels into one state for the complete system. This allows us to significantly enhance the detection time of dangerous operating states, compared to the use of dedicated single sensors.

Sensor and information fusion is recently a major topic which becomes important in machine diagnosis and conditioning for complex production machines and process engineering. It is a known fact that distributed automation systems have a major impact on signal processing and pattern recognition for machine diagnosis. Therefore, it is necessary to research and develop smart diagnosis methods which are applicable for distributed systems like resource-limited cyber-physical systems. In this paper we propose an new approach for sensor and information fusion based on Evidence Theory and socio-psychological decision-making. We show that context based condition monitoring is instantiated even in conflict situations, oc-curing in real life scenarios permanently. A simple but effective importance measure is proposed which controls the significance of conditioning propositions in a system.

Synthetic surfaces, in particular polymer structures, which are used for electronic components, have to be inspected in industrial processes. Polymers show some specific surface characteristics. This a-priori knowledge is useable for the feature extraction of a surface texture and a following classification. The feature extraction is performed by using statistical information, calculated from sum and difference histograms, while the classification is executed by a fuzzy pattern classifier. A defect area can be recognized just on the basis of the tested image and without the need of any further reference learning data. The classification of a defect part is achieved by analyzing the divergence of the extracted feature values from their median related to the inspected area. Surfaces that contain an inconsistent texture will be rejected.

Bank note inspection is a complex task. As more and more print techniques and new security features are established, total quality security and bank note printing must be assured. Therefore, this factor necessitates change of a sensorial concept in general. We propose an optical-acoustical inspection method based upon the concepts of information fusion and fuzzy interpretation of data measures. Furthermore, we present a simplified scheme for information fusion for pattern recognition and data classification based on parametrical unimodal potential functions and a Sugeno-type score value analysis.

Nonlinear spatial transforms and fuzzy pattern classification with unimodal potential functions are established in signal processing. They have proved to be excellent tools in feature extraction and classification. In this paper we present a hardware accelerated image processing and classification scheme for rotation and translation tolerant two-dimensional pattern recognition, which is based on one-dimensional nonlinear discrete circular transforms. However, the scheme is simple; it is stable and therefore well suited for industrial applications. An implementation on one field programmable gate array (FPGA) is proposed.