@uni-wuppertal.de
School of Electrical, Information and Media Engineering
Bergische Universitaet Wuppertal
Artificial Intelligence, Computer Vision and Pattern Recognition, Information Systems, Computer Science
Scopus Publications
Scholar Citations
Scholar h-index
Scholar i10-index
Miguel Alves Gomes, Richard Meyes, Philipp Meisen, and Tobias Meisen
MDPI AG
Alongside natural language processing and computer vision, large learning models have found their way into e-commerce. Especially, for recommender systems and click-through rate prediction, these models have shown great predictive power. In this work, we aim to predict the probability that a customer will click on a given recommendation, given only its current session. Therefore, we propose a two-stage approach consisting of a customer behavior-embedding representation and a recurrent neural network. In the first stage, we train a self-supervised skip-gram embedding on customer activity data. The resulting embedding representation is used in the second stage to encode the customer sequences which are then used as input to the learning model. Our proposed approach diverges from the prevailing trend of utilizing extensive end-to-end models for click-through rate prediction. The experiments, which incorporate a real-world industrial use case and a widely used as well as openly available benchmark dataset, demonstrate that our approach outperforms the current state-of-the-art models. Our approach predicts customers’ click intention with an average F1 accuracy of 94% for the industrial use case which is one percentage point higher than the state-of-the-art baseline and an average F1 accuracy of 79% for the benchmark dataset, which outperforms the best tested state-of-the-art baseline by more than seven percentage points. The results show that, contrary to current trends in that field, large end-to-end models are not always needed. The analysis of our experiments suggests that the reason for the performance of our approach is the self-supervised pre-trained embedding of customer behavior that we use as the customer representation.
Nils Hütten, Miguel Alves Gomes, Florian Hölken, Karlo Andricevic, Richard Meyes, and Tobias Meisen
MDPI AG
Quality assessment in industrial applications is often carried out through visual inspection, usually performed or supported by human domain experts. However, the manual visual inspection of processes and products is error-prone and expensive. It is therefore not surprising that the automation of visual inspection in manufacturing and maintenance is heavily researched and discussed. The use of artificial intelligence as an approach to visual inspection in industrial applications has been considered for decades. Recent successes, driven by advances in deep learning, present a possible paradigm shift and have the potential to facilitate automated visual inspection, even under complex environmental conditions. For this reason, we explore the question of to what extent deep learning is already being used in the field of automated visual inspection and which potential improvements to the state of the art could be realized utilizing concepts from academic research. By conducting an extensive review of the openly accessible literature, we provide an overview of proposed and in-use deep-learning models presented in recent years. Our survey consists of 196 open-access publications, of which 31.7% are manufacturing use cases and 68.3% are maintenance use cases. Furthermore, the survey also shows that the majority of the models currently in use are based on convolutional neural networks, the current de facto standard for image classification, object recognition, or object segmentation tasks. Nevertheless, we see the emergence of vision transformer models that seem to outperform convolutional neural networks but require more resources, which also opens up new research opportunities for the future. Another finding is that in 97% of the publications, the authors use supervised learning techniques to train their models. However, with the median dataset size consisting of 2500 samples, deep-learning models cannot be trained from scratch, so it would be beneficial to use other training paradigms, such as self-supervised learning. In addition, we identified a gap of approximately three years between approaches from deep-learning-based computer vision being published and their introduction in industrial visual inspection applications. Based on our findings, we additionally discuss potential future developments in the area of automated visual inspection.
Tristan Langer, Richard Meyes, and Tobias Meisen
Wiley
AbstractIn recent years, digitization in the industrial sector has increased steadily. Digital data not only allows us to monitor the underlying production process using machine learning methods (anomaly detection, behaviour analysis) but also to understand the underlying production process. Insights from Exploratory Data Analysis (EDA) play an important role in building data‐driven processes because data scientists learn essential characteristics of the data in the context of the domain. Due to the complexity of production processes, it is usually difficult for data scientists to acquire this knowledge by themselves. Hence, they have to rely on continuous close collaboration with domain experts and their acquired domain expertise. However, direct communication does not promote documentation of the knowledge transfer from domain experts to data scientists. In this respect, changing team constellations, for example due to a change in personnel, result in a renewed high level of effort despite the same knowledge transfer problem. As a result, EDA is a cost‐intensive iterative process. We, therefore, investigate a system to extract information from the interactions that domain experts perform during EDA. Our approach relies on recording interactions and system states of an exploration tool and generating guided exploration sessions for domain novices. We implement our approach in a software tool and demonstrate its capabilities using two real‐world use cases from the manufacturing industry. We evaluate its feasibility in a user study to investigate whether domain novices can reproduce the most important insights from domain experts about the datasets of the use cases based on generated EDA sessions. From the results of this study, we conclude the feasibility of our system as participants are able to reproduce on average 86.5% of insights from domain experts.
Basile Tousside, Jörg Frochte, and Tobias Meisen
SCITEPRESS - Science and Technology Publications
: Learning multiple sequentially arriving tasks without forgetting previous knowledge, known as Continual Learning (CL), remains a long-standing challenge for neural networks. Most existing CL methods rely on data replay. However, they are not applicable when past data is unavailable or is not allowed to be synthetically generated. To address this challenge, we propose Sparification and Expansion-based Continual Learning (SECL). SECL avoids forgetting of previous tasks by ensuring the stability of the CNN via a stability regularization term, which prevents filters detected as important for past tasks to deviate too much when learning a new task. On top of that, SECL makes the network plastic via a plasticity regularization term that leverage the over-parameterization of CNNs to efficiently sparsify the network and tunes unimportant filters making them relevant for future tasks. Also, SECL enhances the plasticity of the network through a simple but effective heuristic mechanism that automatically decides when and where (at which layers) to expand the network. Experiments on popular CL vision benchmarks show that SECL leads to significant improvements over state-of-the-art method in terms of overall CL performance, as measured by classification accuracy as well as in terms of avoiding catastrophic forgetting.
Vladimir Samsonov, Enslin Chrismarie, Hans-Georg Köpken, Schirin Bär, Daniel Lütticke, and Tobias Meisen
Springer Science and Business Media LLC
AbstractComputer Numerical Control (CNC) milling is a commonly used manufacturing process with a high level of automation. Nevertheless, setting up a new CNC milling process involves multiple development steps relying heavily on human expertise. In this work, we focus on positioning and orientation of the workpiece (WP) in the working space of a CNC milling machine and propose a deep learning approach to speed up this process significantly. The selection of the WP’s setup depends on the chosen milling technological process, the geometry of the WP, and the capabilities of the considered CNC machining. It directly impacts the milling quality, machine wear, and overall energy consumption. Our approach relies on representation learning of the milling technological process with the subsequent use of reinforcement learning (RL) for the WP positioning and orientation. Solutions proposed by the RL agent are used as a warm start for simple hill-climbing heuristics, which boosts overall performance while keeping the overall number of search iterations low. The novelty of the developed approach is the ability to conduct the WP setup optimization covering both WP positioning and orientation while ensuring the axis collision avoidance, minimization of the axis traveled distances and improving the dynamic characteristics of the milling process with no input from human experts. Experiments show the potential of the proposed learning-based approach to generate almost comparably good WP setups order of magnitude faster than common metaheuristics, such as genetic algorithms (GA) and Particle Swarm Optimisation (PSA).
Friedrich von Bülow, Markus Wassermann, and Tobias Meisen
Elsevier BV
Miguel Alves Gomes, Mark Wönkhaus, Philipp Meisen, and Tobias Meisen
MDPI AG
Real-time customer purchase prediction tries to predict which products a customer will buy next. Depending on the approach used, this involves using data such as the customer’s past purchases, his or her search queries, the time spent on a product page, the customer’s age and gender, and other demographic information. These predictions are then used to generate personalized recommendations and offers for the customer. A variety of approaches already exist for real-time customer purchase prediction. However, these typically require expertise to create customer representations. Recently, embedding-based approaches have shown that customer representations can be effectively learned. In this regard, however, the current state-of-the-art does not consider activity time. In this work, we propose an extended embedding approach to represent the customer behavior of a session for both known and unknown customers by including the activity time. We train a long short-term memory with our representation. We show with empirical experiments on three different real-world datasets that encoding activity time into the embedding increases the performance of the prediction and outperforms the current approaches used.
Johannes Benkert, Robert Maack, and Tobias Meisen
MDPI AG
In recent years, a steady increase in maritime business and annual container throughput has been recorded. To meet this growing demand, terminal operators worldwide are turning to automated container handling. For the automated operation of a crane, a reliable capture of the environment is required. In current state-of-the-art applications this is mostly achieved with light detection and ranging (LiDAR) sensors. These sensors enable precise three-dimensional sampling of the surroundings, even at great distances. However, the use of LiDAR sensors has a number of disadvantages, such as high acquisition costs and limited mounting positions. This raises the question of whether the LiDAR systems of automated container terminals (ACT) can be replaced with cameras. However, this transformation is not easy to accomplish and is explored in more depth in this paper. The field of camera-based container automation presented in this publication is largely unexplored. To the best of our knowledge, there is currently no automated container terminal in real-world operation that exclusively uses cameras. This publication aims to create a basis for further scientific research towards the goal of a fully camera-based container automation. Therefore, the authors present a narrative review providing a broad overview of the mentioned transformation, identifying research gaps, and suggesting areas for future research. In order to achieve this, this publication examines the fundamentals of an automated container terminal, the existing automation solutions and sensor technologies, as well as the opportunities and challenges of a transformation from LiDAR to camera.
Miguel Alves Gomes and Tobias Meisen
Springer Science and Business Media LLC
AbstractThe importance of customer-oriented marketing has increased for companies in recent decades. With the advent of one-customer strategies, especially in e-commerce, traditional mass marketing in this area is becoming increasingly obsolete as customer-specific targeting becomes realizable. Such a strategy makes it essential to develop an underlying understanding of the interests and motivations of the individual customer. One method frequently used for this purpose is segmentation, which has evolved steadily in recent years. The aim of this paper is to provide a structured overview of the different segmentation methods and their current state of the art. For this purpose, we conducted an extensive literature search in which 105 publications between the years 2000 and 2022 were identified that deal with the analysis of customer behavior using segmentation methods. Based on this paper corpus, we provide a comprehensive review of the used methods. In addition, we examine the applied methods for temporal trends and for their applicability to different data set dimensionalities. Based on this paper corpus, we identified a four-phase process consisting of information (data) collection, customer representation, customer analysis via segmentation and customer targeting. With respect to customer representation and customer analysis by segmentation, we provide a comprehensive overview of the methods used in these process steps. We also take a look at temporal trends and the applicability to different dataset dimensionalities. In summary, customer representation is mainly solved by manual feature selection or RFM analysis. The most commonly used segmentation method is k-means, regardless of the use case and the amount of data. It is interesting to note that it has been widely used in recent years.
Hasan Tercan and Tobias Meisen
ACM
The digitization of manufacturing processes opens up the possibility of using machine learning methods on process data to predict future product quality. Based on the model predictions, quality improvement actions can be taken at an early stage. However, significant challenges must be overcome to successfully implement the predictions. Production lines are subject to hardware and memory limitations and are characterized by constant changes in quality influencing factors. In this paper, we address these challenges and present an online prediction approach for real-world manufacturing processes. On the one hand, it includes methods for feature extraction and selection from multimodal process and sensor data. On the other hand, a continual learning method based on memory-aware synapses is developed to efficiently train an artificial neural network over process changes. We deploy and evaluate the approach in a windshield production process. Our experimental evaluation shows that the model can accurately predict windshield quality and achieve significant process improvement. By comparing with other learning strategies such as transfer learning, we also show that the continual learning method both prevents catastrophic forgetting of the model and maintains its data efficiency.
Rebecca Braken, Alexander Paulus, André Pomp, and Tobias Meisen
MDPI AG
Semantic models are utilized to add context information to datasets and make data accessible and understandable in applications such as dataspaces. Since the creation of such models is a time-consuming task that has to be performed by a human expert, different approaches to automate or support this process exist. A recurring problem is the task of link prediction, i.e., the automatic prediction of links between nodes in a graph, in this case semantic models, usually based on machine learning techniques. While, in general, semantic models are trained and evaluated on large reference datasets, these conditions often do not match the domain-specific real-world applications wherein only a small amount of existing data is available (the cold-start problem). In this study, we evaluated the performance of link prediction algorithms when datasets of a smaller size were used for training (few-shot scenarios). Based on the reported performance evaluation, we first selected algorithms for link prediction and then evaluated the performance of the selected subset using multiple reduced datasets. The results showed that two of the three selected algorithms were suitable for the task of link prediction in few-shot scenarios.
Constantin Waubert de Puiseau, Jannik Peters, Christian Dörpelkus, Hasan Tercan, and Tobias Meisen
Elsevier BV
Alexander Paulus, André Pomp, and Tobias Meisen
ACM
Modern data management is evolving from centralized integration-based solutions to a non-integration-based process of finding, accessing and processing data, as observed within dataspaces. Common reference dataspace architectures assume that sources publish their own domain-specific schema. These schemas, also known as semantic models, can only be partially created automatically and require oversight and refinement by human modellers. Non-expert users, such as mechanical engineers or municipal workers, often have difficulty building models because they are faced with multiple ontologies, classes, and relations, and existing tools are not designed for non-expert users. The PLASMA framework consists of a platform and auxiliary services that focus on providing non-expert users with an accessible way to create and edit semantic models, combining automation approaches and support systems such as a recommendation engine. It also provides data conversion from raw data to RDF. In this paper we highlight the main features, like the modeling interface and the data conversion engine. We discuss how PLASMA as a tool is suitable for building semantic models by non-expert users in the context of dataspaces and show some applications where PLASMA has already been used in data management projects.
André Pomp, Maike Jansen, Holger Berg, and Tobias Meisen
ACM
The circular economy (CE) is essential to achieving a sustainable future through resource conservation and climate protection. Efficient use of materials and products over time is a critical aspect of CE, helping to reduce CO2 emissions, waste and resource consumption. The Digital Product Passport (DPP) is a CE-specific approach that contains information about components and their origin, and can also provide environmental and social impact assessments. However, creating a DPP requires collecting and analyzing data from many different stakeholders along the supply chain and even throughout the product lifecycle. In this paper, we present a concept for the SPACE_DS, which is a data space for circular economy data. A key point here is that the SPACE_DS enables the creation of DPPs by especially considering privacy and security concerns of data providers.
Tristan Langer, André Pomp, and Tobias Meisen
ACM
Capturing, visualizing and analyzing provenance data to better understand and support analytic reasoning processes is a rapidly growing research field named analytic provenance. Provenance data includes the state of a visualization within a tool as well as the user’s interactions performed while interacting with the tool. Research in this field has produced in many new approaches that generate data for specific tools and use cases. However, since a variety of tools are used and analytic tasks are performed in real analysis use cases there is a problem in building an interoperable baseline data corpus for investigation of the transferability of different approaches. In this paper, we present a visionary data space architecture for integrating and processing analytic provenance data in a unified way using semantic modeling. We discuss emerging challenges and research opportunities to realize such a vision using semantic models in data spaces to enable analytic provenance data interoperability.
Maike Jansen, Tobias Meisen, Christiane Plociennik, Holger Berg, André Pomp, and Waldemar Windholz
MDPI AG
The Digital Product Passport (DPP) is a concept for collecting and sharing product-related information along the life cycle of a product. DPPs are currently the subject of intense discussion, and various development efforts are being undertaken. These are supported by regulatory activities, especially in the case of the battery passport. The aggregation of product life-cycle data and their respective use, as well as the sharing of these data between companies, entrepreneurs, and other actors in the value chain, is crucial for the creation of a resource-efficient circular economy. Despite the urgent need for such a solution, there is currently little attention given to the digital infrastructure for the creation and handling of the DPPs (i.e., the so-called DPP system). Moreover, there is so far no common understanding of what the requirements for a DPP system are. This is the background and underlying motivation of our paper: we identify the requirements for a DPP system in a structured way, i.e., based on stakeholder involvement and current literature from science and industry. In addition, we compose, categorize, and critically analyze the results, i.e., the list of requirements for DPP systems, in order to identify gaps. Summarized, our research provides insights into the criteria to be considered in the creation of an actual DPP system.
Yannik Hahn, Tristan Langer, Richard Meyes, and Tobias Meisen
MDPI AG
Deep learning models have revolutionized research fields like computer vision and natural language processing by outperforming traditional models in multiple tasks. However, the field of time series analysis, especially time series forecasting, has not seen a similar revolution, despite forecasting being one of the most prominent tasks of predictive data analytics. One crucial problem for time series forecasting is the lack of large, domain-independent benchmark datasets and a competitive research environment, e.g., annual large-scale challenges, that would spur the development of new models, as was the case for CV and NLP. Furthermore, the focus of time series forecasting research is primarily domain-driven, resulting in many highly individual and domain-specific datasets. Consequently, the progress in the entire field is slowed down due to a lack of comparability across models trained on a single benchmark dataset and on a variety of different forecasting challenges. In this paper, we first explore this problem in more detail and derive the need for a comprehensive, domain-unspecific overview of the state-of-the-art of commonly used datasets for prediction tasks. In doing so, we provide an overview of these datasets and improve comparability in time series forecasting by introducing a method to find similar datasets which can be utilized to test a newly developed model. Ultimately, our survey paves the way towards developing a single widely used and accepted benchmark dataset for time series data, built on the various frequently used datasets surveyed in this paper.
Yannik Hahn, Robert Maack, Guido Buchholz, Marion Purrio, Matthias Angerhausen, Hasan Tercan, and Tobias Meisen
Elsevier BV
Mario Angos Mediavilla, Michele Lagnese, André Pomp, and Tobias Meisen
Elsevier BV
Christian Bitter, Jannik Peters, Hasan Tercan, and Tobias Meisen
Elsevier BV
Yannik Steiniger, Angel Bueno, Dieter Kraus, and Tobias Meisen
IEEE
Data scarcity remains the main challenge when developing deep learning models for sonar image analysis. Although dataset augmentation with synthetically generated images has been proposed, these methods are far from optimal as they are unable to capture the range of physical factors affecting sonar images, given the small data regimes used for their training. This work focuses on an alternative solution and investigates the learning of suitable representations for classifying small-sized sonar datasets. To achieve this, we propose a new approach that entails the combination of convolutional and scattering neural networks, a wavelet-based neural network that produces feature map representations robust to image variations. Our experiments show that these representations are easier to classify, leading to a performance increase of 4.5 percentage points in F1-score for the combined network compared to a plain convolutional neural network. Furthermore, we interpret the representation obtained by the scattering transformation as robust feature descriptors, where the geometric shapes of underwater objects are rendered prominent and stable to minor sonar distortions.
Tristan Langer, Viktor Welbers, Yannik Hahn, Mark Wönkhaus, Richard Meyes, and Tobias Meisen
Springer Nature Switzerland