Search

The reliable detection and precise assessment of white etching layers (WEL) are key challenges in the investigation of a component's surface integrity. This paper proposes an innovative methodology for evaluating the extent of WEL in quenched and tempered steels, machined by Boring and Trepanning Association (BTA) deep hole drilling. Micrographs obtained by light microscopy were partitioned into classes by three methods, separating the WEL from the base material and the embedding resin. Traditional manual segmentation was performed as a benchmark for automatic segmentation methods. A gray level thresholding-based method served for the segmentation of micrographs partitioned into subsets. In addition to conventional manual and thresholding-based segmentation, a machine learning-based approach for image segmentation was applied. The segmented images were further analyzed by a newly developed set of algorithms, implemented to obtain detailed information on the WEL, e.g. their average thickness as well as the area covered by WEL in the micrographs. Results indicate that both, gray level thresholding, as well as machine learning-based image segmentation, show potential for the automated diagnosis and assessment of WEL. They both yield quantitatively similar, but less biased results compared to manual segmentation. [ABSTRACT FROM AUTHOR]

The accurate estimation of 6D pose remains a challenging task within the computer vision domain, even when utilizing 3D point cloud data. Conversely, in the manufacturing domain, instances arise where leveraging prior knowledge can yield advancements in this endeavor. This study focuses on the disassembly of starter motors to augment the engineering of product life cycles. A pivotal objective in this context involves the identification and 6D pose estimation of bolts affixed to the motors, facilitating automated disassembly within the manufacturing workflow. Complicating matters, the presence of occlusions and the limitations of single-view data acquisition, notably when motors are placed in a clamping system, obscure certain portions and render some bolts imperceptible. Consequently, the development of a comprehensive pipeline capable of acquiring complete bolt information is imperative to avoid oversight in bolt detection. In this paper, employing the task of bolt detection within the scope of our project as a pertinent use case, we introduce a meticulously devised pipeline. This multi-stage pipeline effectively captures the 6D information with regard to all bolts on the motor, thereby showcasing the effective utilization of prior knowledge in handling this challenging task. The proposed methodology not only contributes to the field of 6D pose estimation but also underscores the viability of integrating domain-specific insights to tackle complex problems in manufacturing and automation. [ABSTRACT FROM AUTHOR]

The performance of a manufacturing system is a result of the interrelated characteristics of all functional units within the system, including material flow. Therefore, the operation of a manufacturing system requires harmonization of its elements, resulting in an optimized material flow, reduced operating costs, and energy consumption. Existing methods approach this problem by applying analytic formulations or material flow simulations that represent the behavior of a manufacturing system. However, building such a model is time-consuming and requires making assumptions and simplifications, which can lead to inaccurate results. As a consequence, this paper proposes two novel machine learning (ML) models that are trained to represent the characteristics of manufacturing systems. The first model is trained to predict the properties of the material flow and the functional units used for production. The second model predicts whether the manufacturing system is in a regular state or whether anomalies are present. The capabilities of the ML models are investigated with three application scenarios of increasing complexity which are used to generate synthetic training data generated in a discrete event simulation. The results show that the ML models can describe the characteristics of a manufacturing system with high accuracy which offers various possibilities for future research and application. [ABSTRACT FROM AUTHOR]

The efficient and safe movement of goods and people require reliable railway systems. Quality assurance of manufactured and assembled systems and correct maintenance of such systems are required to keep rolling stock in good operational condition. Quality assurance and maintenance in the railway industry can be costly and time-consuming, but the expansive growth of data due to smart sensors and monitoring technologies makes it possible to leverage the potential of machine learning to reduce cost and labour. Improved reliability and safety, and reduced costs are benefits that the use of "Big Data" and machine learning techniques can realise. However, despite these potential benefits for manufacturers, rail operators, and passengers, the rail industry is still labelled for its lack of innovation, while in most other industries, data is regarded as a strategic asset for competitive advantage. This paper demonstrates how machine learning and data analysis can be used to benefit railway industry manufacturers and operators when applied to rolling stock data. It also illustrates the lost opportunity in the rail industry for not applying data-driven solutions to their full potential. The paper also discusses the current applications of machine learning in the railway industry and provides the requirements for the implementation of machine learning techniques. Machine learning is applied to pantograph data of a South African railway operator's rolling stock. Classification – a machine learning technique – is used to identify and categorise events within the dataset to discover whether pantograph bounce occurs due to faulty sensors, faulty pantographs, or defective infrastructure. In this paper it is demonstrated how machine learning can benefit rail manufacturers and operators to improve manufacturing and assembly processes, as well as maintenance practices. It is concluded that railways should treat data similarly to other railway assets, with suitable management and governance practices. [ABSTRACT FROM AUTHOR]

This paper studies the challenges encountered during the implementation of machine learning (ML) applications in production planning and control (PPC) functions of manufacturing companies. The application of ML in PPC promises numerous benefits, and several studies have revealed the positive impact of such an implementation on key performance indicators of production systems, such as lead time, due date reliability, and inventory levels. However, despite the theoretical potential and the increased research interest in the field over the last five years, practical applications remain rare and discourse on the existence of this gap between research and practice remains relatively sparse. In this paper, we identify general ML implementation barriers in the manufacturing domain based on a structured literature review and evaluate the relevance of these barriers within the PPC domain by interviewing ML and PPC experts. This research is beneficial for practitioners and researchers who aim to implement ML applications in PPC by enabling them to evaluate the possible barriers and challenges during implementation projects within their organizations. [ABSTRACT FROM AUTHOR]

Kanban systems are still the most popular method for pull control of material replenishment. In the Industry 4.0 era, the so-called eKanban represents a digital version of the classic Kanban. Thereby, in some eKanban variants, capturing actual stock-level by digital sensors would help or helps already to reduce capturing time and human failures. Currently, such digital sensors are not really beneficial because they are wired and expensive to install. In this paper, we investigate the potential of battery-powered stock-level sensors for eKanban systems using a simulation study. Battery- powered stock-level sensors are easier to install and, therefore, more beneficial than wired sensors if the battery´s lifetime is long enough. Because of the limited energy, it makes sense to trigger such sensor systems dynamically depending on the actual and predicted stock levels and on order points for minimum stock. For example, the sensor systems usually sleep and capture the stock-level only in a fixed interval for example every hour. Besides fixed intervals, it can be useful to trigger such sensors dynamically by reducing the activations in time with no or few material take-outs. To identify the best trigger strategy, this paper investigates how artificial intelligence (AI) methods, in this case, reinforcement learning, can be deployed. Thereby, the used material flow simulation maps the application scenario of a four-stage assembly line and compares the AI-based trigger with a manually operated demand notification. In this way, it can be shown that battery lifetime is increased fourteen times more compared to the manually operated demand notification, and battery runtimes of several months can be achieved by AI-based optimization of the sensor energy consumption. This approach leads to an eKanban solution becoming more economical, and thus smaller companies can also consider digitizing their material flow control [ABSTRACT FROM AUTHOR]

The tool lifetime and the material removal rate significantly influence the production costs of a machined component. In order to maintain process reliability, tools are often changed too early in industrial practice, which is a waste of resources and results in increased tool costs and extended setup times. Tool condition monitoring systems based on machine learning have shown great potential to detect and predict tool wear, thereby reducing the risk of tool failure and optimising tool change intervals. Nevertheless, common supervised learning methods for tool wear detection require extensive data generation and high labeling effort. This paper presents two approaches with minimal to no labeling effort for predicting the end of the service life of drilling tools based on anomaly detection. Therefore, a classification approach to anomaly detection based on XGBoost and an unsupervised autoencoder architecture are evaluated on data acquired during machining experiments. In addition, the performance when applying the models of both approaches to other cutting parameters is investigated. Beyond a high classification performance on all datasets, the XGBoost approach can also be used as to derive an indicator for reaching the end of life, even when applied to other cutting parameters. For the autoencoder approach, it is shown that its reconstruction error correlates with the anomaly label and thus can also be used for anomaly detection. [ABSTRACT FROM AUTHOR]

Production system simulation is a powerful tool for optimizing the use of resources on both the planning and control level. However, creating and tuning such models manually is a tedious and error-prone task. Despite some approaches to automate this process, the state-of-the-art relies on the generation of models, by incorporating the knowledge of experts. Nevertheless, effectively creating and tuning such production simulations is, thus, a key driver for reducing costs, carbon footprint, and tardiness and therefore an essential factor in today´s production. Beneficial would be automated and flexible frameworks, since these are applicable to different use cases requiring less effort. Yet, in the age of Industry 4.0, data is ubiquitous and easily available and can serve as a basis for virtual models representing reality. Increasingly, these virtual models shall be interlinked with the current state of real-world systems to form so-called digital twins. As automated and flexible frameworks are missing, this paper proposes a novel approach where observed real system behavior is used and fed into a large-scale machine learning model trained on a plethora of possible parameter sets. The main target is to train this machine learning model to minimize the reality gap between the behavior of the simulated and real system by selecting corresponding simulation system parameters. By estimating those parameters an enhancement of the simulation will emerge. An interlink to real systems can be derived resulting in a digital shadow which is capable to forecast the future similarly to reality. The approach to overcoming the gap between reality and simulation (real2sim) is validated in simulations. [ABSTRACT FROM AUTHOR]

This paper presents a comprehensive literature review of the research and application of machine learning (ML) algorithms in recommender systems (RS). The study aims to identify recent trends, explore real-life applications, and guide researchers in positioning their research activities in this domain published in 2023 (Jan-June). The findings are categorized into different domains including education, healthcare, ML algorithms (auto-encoders and reinforcement learning), e-commerce, and digital journalism. The review highlights the enhanced recommendation accuracy, increased scalability, personalization and context awareness, diverse ML techniques, and strategies for handling cold start and data sparsity, and the foundation for future advancements in ML algorithms for RSs considering the application in manufacturing enterprises. [ABSTRACT FROM AUTHOR]

This paper describes an artificial intelligence (AI) based failure management approach across the value chain for the commercial vehicle industry by integrating and utilizing lifecycle data for product and production optimization. The amount of available data throughout a product lifecycle has increased significantly in previous years, primarily driven by the development and deployment of cyber-physical systems. While data from a single entity in the value chain already enables failure management-related analysis and services, including AI-based methods such as predictive maintenance, there remains a lack of systematic approaches to utilize data across the entire value chain. This paper proposes an AI-based failure management approach, which relies on integrating a variety of diverse data sources along the value chain. At first, three so-called application areas were defined: process and product optimization, availability optimization, and performance optimization. Consequently, practice-relevant use cases are identified for each area, for which it is shown how failures in the value chain can be proactively eliminated with the support of AI. Methods for predictive analytics are adapted for cross-value chain failure management to derive correlations between different stages of the production process and product usage. Based on these results and human expert knowledge, proactive measures are recommended by a decision support system (DSS) to resolve failures before arising. The commercial vehicle industry serves as an overarching validation case study for the practice-relevant verification of the targeted applications. The paper gives an outlook on the envisaged research work for the realization of holistic failure management. [ABSTRACT FROM AUTHOR]

Anomaly detection is a key feature to monitor production systems an avoid downtimes, which has still not been implemented holistically. Especially in highly flexible plants or special process machines, conventional approaches like neuronal network classification or intelligent autoencoder fault detection are not suitable firstly due to the small amount and secondly due to the lack of labeling of data for each process. In this paper a novel concept is presented to segment different processes intelligently in the first step to find fine granular process patterns across process boundaries. Based on these patterns, anomaly detection and further classification are performed. A special feature is the integration of user knowledge, so that classification is possible even with a small amount of data. This approach is validated on an assembly line for electric motor production as well as in a handling robot. This paper shows results from real test series and thus demonstrates the practical suitability of the novel approach. [ABSTRACT FROM AUTHOR]

Factory layout planning is a recurring and important process since an inadequately planned layout can significantly impede the operation of a factory. A time-efficient and automated planning method is especially important as shortened product lifecycles, changing technologies, and market requirements lead to changes in the production program, which further require the re-planning of the factory layout. Reinforcement learning (RL) has proven to be a suitable approach to support decision-making in complex layout planning problems. However, recent studies only consider comparably small and exemplary problem statements with a maximum of 4 functional units. In consequence, these approaches are insufficient for industrial applications in which numerous functional units are considered. Regarding this background, this paper investigates the scalability of a RL-based planning approach, in which a Double Deep Q Learning agent is used to solve a layout planning task for a small batch production with 20 functional units. Considering that recent approaches lack scalability, this paper presents a new state representation in combination with an action masking method that optimizes the action selection process to ensure the scalability of the approach and reduce the training time. After the training, the algorithm can generate an optimized factory layout by reducing the resulting transportation time for industrial relevant problem sizes. [ABSTRACT FROM AUTHOR]

Manufacturing processes have undergone tremendous technological progress in recent decades. To meet the agile philosophy in industry, data-driven algorithms need to handle growing complexity, particularly in Computer Numerical Control machining. To enhance the scalability of machine learning in real-world applications, this paper presents a benchmark dataset for process monitoring of brownfield milling machines based on acceleration data. The data is collected from a real-world production plant using a smart data collection system over a two-years period. In this work, the edge-to-cloud setup is presented followed by an extensive description of the different normal and abnormal processes. An analysis of the dataset highlights the challenges of machine learning in industry caused by the environmental and industrial factors. The new dataset is published with this paper and available at: https://github.com/boschresearch/CNC_Machining. [ABSTRACT FROM AUTHOR]

Due to the high number of influences in production applications, quality fluctuation and deviating results in thermal spray applications are a challenging issue in manufacturing processes. This paper presents a concept for an adaptive process control based on different machine learning algorithms that are utilized to correlate input parameters with output results in thermal spray processes. For that purpose, a test environment was developed, incorporating different application scenarios with various input parameters. During the experiments, a series of measurements was performed, that is used to describe and evaluate the process. The extracted data of the conducted experiments are transformed into features and serve as the data input for the development of machine learning models based on various algorithms. These algorithms offer a prediction of the peak position, which describes the deviation of the programmed application spot, thus offering a quality prediction by utilizing solely the input process parameters. This work provides an approach based on the compared algorithms, that can be applied to create an adaptive process control, by adjusting the input parameters during the thermal spray application process. [ABSTRACT FROM AUTHOR]

Since bolted joints are ubiquitous in manufacturing, their effective and reliable quality assurance is particularly important. Most tightening processes rely on statistical methods to detect faulty screw connections already during assembly. In this paper, we address the detection of faulty tightening processes using a clustering based approach from the field of Unsupervised Machine Learning. In particular, we deploy the k -Means algorithm on a real-world dataset from the automotive industry. The model uses Dynamic Time Warping to determine the similarity between the normal and abnormal tightening processes, treating each one as an independent temporal sequence. This approach offers three advantages compared to existing supervised methods: 1.) time series with different lengths can be utilized without extensive preprocessing steps, 2.) errors never seen before can be detected using the unsupervised approach, and 3.) extensive manual efforts to generate labels are no longer necessary. To evaluate the approach, it is applied in a scenario where actual class labels are available. This allows evaluating the clustering results using traditional classification scores. The approach manages to achieve an accuracy of up to 88.89% and a macro-average F1-score of up to 63.65%. [ABSTRACT FROM AUTHOR]

Companies with Make-to-Order (MTO) manufacturing have always faced the conflict of meeting a large volume of individual customer orders on time while remaining as flexible as possible. Unlike Make-to-Stock (MTS) manufacturing, planning the production requirements for MTO is a challenging task since incoming orders may vary in time and quantity, while also being subject to a number of variables. In some cases, the delivery time allocated by the customer can be less than the required Order Lead Time to fulfil the order. Manufacturers can respond to this with either with approached from production management or from data science. This paper presents a framework to leverage the benefit of both domains. We conduct a literature review and present the results of an expert workshop. We propose criteria for a suitable data enrichment and the application of Machine Learning (ML) methods in sales and demand forecasting. In conclusion, the proposed framework helps to equip manufacturing companies with a structured strategy for data management and to utilize the benefit of ML for sales and demand forecasting. [ABSTRACT FROM AUTHOR]

In the perspective of feed-forward control through a manufacturing process chain, production data can be used downstream to correct subsequent processes and improve product quality. In an extrusion blow-moulding (EBM) use case, supervised learning (SL) has been applied to predict geometrical dimensions based on process data, as a possible basis for feed-forward control. The labeled tabular dataset showed signifcant time dependency which complicates the learning task. In this paper we present a time-series regression approach and compare three common SL algorithms – Random Forest, Gradient Boosting and XGBoost. The results show that only part of the variations in product geometry could be learned from process data and that future work will be necessary in order to increase the models' performance. [ABSTRACT FROM AUTHOR]

Studies affecting manufacturing system resilience have been of rising interest in recent years due to increased rates of naturally caused, and socio-economic disturbances. The non-linearity in modern manufacturing systems, attributed to a host of internal and external complexities in structure and operations cannot be properly modeled using conventional white-box techniques. This paper focuses on building a black-box description of a manufacturing system as a solution, addressing the inherent non-linearity, and further supporting fast investigations on a larger solution space with the example of an application scenario. This work acts as a precursor towards the implementation of a conceptual workflow describing a cross-domain application of machine learning (ML) for optimizing manufacturing system resilience. [ABSTRACT FROM AUTHOR]

The production of lithium-ion batteries is characterized by a high complexity and manifold cause-effect correlations between process parameters and product properties, especially in electrode production processes. Semi-dry electrode production for lithium-ion batteries allows a reduction in energy consumption during production due to a reduction of drying length and decreases the amounts of toxic solvents. In semi-dry electrode production, highly viscous granules are processed using a two-roller calender in contrast to a low-viscous slurry in conventional coating processes. For this two-roller calender applied in semi-dry electrode production, the relationships between process and product parameters are unknown and related processes, e.g., the conventional calendering process, have limited validity. This paper presents the approach to identifying cause-effect correlations in semi-dry electrode production using a calender with at least three main process parameter dimensions, such as roller temperature and speed as well as calender gap, leading to a large parameter space. In the first step, the most crucial process parameters in the semi-dry coating step were identified. A concept for setting up the digital infrastructure for data acquisition and controlling the calender during experiments to generate a suitable data base is proposed. In the second step, this data can be used to understand and model the semi-dry electrode manufacturing process and to identify optimized parameters for specific product properties based on machine learning algorithms. A multi-fidelity model in combination with gaussian process regression is a promising way to iteratively adapt the model to new materials or different granule formulations. Also, a straightforward extension based on the presented concept for other calender types is possible. [ABSTRACT FROM AUTHOR]

In the domain of process planning, operation sequencing is one of the most essential tasks. Operation sequences determine the optimal order of different manufacturing processes for a product and therefore influence cost, quality, and production time of the manufactured product. Current approaches predicting operation sequences formulate the task as a mathematical optimization problem or rule-based system and therefore require a formulation of knowledge, for example as precedence constraints. Such a formulation is tedious and complex, because a lot of different relationships have to be considered. Moreover, due to individual machining equipment of companies the knowledge formulation can vary and consequently needs to be adopted to the manufacturing systems specifics. Therefore, this paper presents a deep artificial neural networks (ANN) approach for which a mathematical formulation for the operation sequencing task is not required and the system can quickly be adopted to different manufacturing systems through training processes. The presented approach captures knowledge from historic process plans through training. The deep ANN is constructed combining a convolutional neural network (CNN) with a long short-term memory network (LSTM) to predict operation sequences. The ANN is trained using business data including a wide range of machining operations. [ABSTRACT FROM AUTHOR]

Due to global competition and continuously changing customer demand, manufacturers nowadays face frequent and unpredictable market shifts. Introducing reconfigurability into contemporary manufacturing systems can enhance cost-effective and rapid responsiveness to these variations. A Reconfigurable Manufacturing System (RMS) can provide a tailored production process in response to changes in operating procedures or machine statuses. Just like any other manufacturing system, RMS requires effective and timely diagnosis as well as prognosis to function smoothly. A Reconfigurable Inspection System (RIS) is designed within an RMS for data-oriented detection of product quality with a minimum number of inspection units. Existing studies about reconfiguration, however, focus on production while disregarding inspection. Artificial Intelligence (AI) has the potential to significantly assist manufacturers over the next decade due to their heavy dependency on data. AI applications such as Machine learning (ML) and Deep Learning (DL) can aid in addressing issues such as tracking manufacturing failures back to specific phases in the manufacturing process by learning relevant data patterns. Thus, this paper aims to provide an overview of the current literature on RMS as well as ML/DL technologies that can be integrated into RIS to enhance performance. Subsequently, a comprehensive model of an AI-based RIS is proposed based on the experimental results derived from existing publications, and the retrofitting procedure of a case study is presented. However, the proposed model and the retrofitting procedure are not validated by experimental results or physical implementation. [ABSTRACT FROM AUTHOR]

The integration of a digital twin into inspection planning enables a novel procedure that reduces avoidable inspection times and costs. This paper shows a method for component-specific adaption of inspection plans by feeding back data-based quality results into inspection planning. An initial evaluation of the method on a real aerospace aluminum component is carried out using a 3-axis milling process. Machine learning based quality models were implemented for the inspection features shape deviation and surface roughness. With the knowledge gained, the inspection time for the process can be reduced by up to 75 % per component. [ABSTRACT FROM AUTHOR]

In current product development methods, the development of a system's information processing control unit is done by defining the exact algorithm of "how" to solve a problem. This is a major restriction because it relies on human creativity and problem-solving skills. Additionally and particularly in complex technical systems, problems exist that are intuitive for a human to solve, but the definition of a problem-solving algorithm is not trivial. In this paper, a method is presented and described that addresses the current challenges in the product development process by leveraging the advantages and properties of reinforcement learning. The presented method aims to avoid the complex programming process for the information processing of a system by not providing a specific definition of "how" a problem should be solved, but rather a specific definition of "what" should be solved. Through reinforcement learning-based development, the solution for a defined problem can be generated. In addition, this method is intended to respond to changing conditions, and the ability to make rapid adjustments when already deployed. This reduces the overall complexity during the product development process. The presented method is successfully applied to an exemplary use-case. [ABSTRACT FROM AUTHOR]

Working with information and knowledge is highly important in engineering design. Especially due to the continuous growth of information embedded in documents, the extraction process needs to be automated and accelerated. Hereby, engineering standards are an important source of knowledge in product development. However, the current way of working with these documents is inefficient and requires a lot of manual steps. This research paper focuses on the automatic extraction of requirements from standards documents and aims to systematically evaluate different extraction techniques in order to identify the best suitable technique for the given problem. For this, a dataset with approximately 10,000 entries is generated from existing standards. It is used to evaluate different extraction techniques whereas rule-based as well as supervised and unsupervised machine-learning techniques are implemented. It is shown that requirements are extracted efficiently, i.e. the F1-score is mostly greater than 80%. This proves the overall suitability for automatic requirements extraction from standards documents. The best results are achieved by Support Vector Machine (F1-score SVM =89.6%), followed by the fine-tuned BERT model (F1-score BERT =87.9%). Additionally, the rule-based extraction shows promising results and should be considered in future due to its high transparency. Future work will focus on the optimization and combination of different techniques as well as the automatic integration into requirements engineering tools. [ABSTRACT FROM AUTHOR]

Parts with complex surfaces have become increasingly exploited due to advanced manufacturing processes. The extensive use of these complex parts in industry in recent years has led to new challenges within the design process when simulating their geometric defects for function assessment, quality control and tolerance analysis. Compared to mechanical parts with regular geometric surfaces and well-established geometrical specification standards, the geometry of complex surfaces is often difficult to be represented by explicit parameters or processed by feature operations, which poses challenges to the modelling and simulation of geometric deviations. Skin Model Shapes (SMSs) enable a discrete representation of surface deviations that allows the storage, simulation and analysis of shapes' geometric variations, and have been successfully applied for canonical geometries. In this paper, a method for generating geometric deviations on parts with complex surface based on deep learning is proposed. Laplace-Beltrami Operator (LBO) is exploited to encode geometric deviations as manufacturing signature and deviations patterns. After pre-processing steps, deviation patterns selected from modal decomposition are used as training data for the mesh Convolutional Neural Networks (CNN), which establishes the links between patterns and parts' manufacturing factors, and transferred to target shapes. Geometric deviations are generated on target shapes by linear combination of the transferred patterns with respect to manufacturing signature. A case study is implemented from an open dataset to illustrate the proposed method. The simulated geometric deviations show their effectiveness and potential to improve the design process and the manufacturing accuracy of parts with complex surfaces. [ABSTRACT FROM AUTHOR]

Clamping fixtures are located in the accuracy path of the manufacturing system. Therefore, their dynamic behavior has a direct impact on the accuracy, quality and cost effectiveness of the machined workpieces. However, due to the assembled structures, the nonlinear dynamic contact behavior is significantly related to the stiffness and damping properties of the complete system. In order to precisely describe the nonlinear phenomena, e.g. microslip effects, the simulation requires more realistic friction models than the Coulomb contact law, which is widely used in FE software. This paper focuses on modeling the workpiece-fixture contact behavior based on experimental measurements. The normal contact was modeled using the constitutive model of Greenwood and Williamson. In tangential direction, the Iwan model was applied for the hysteresis loops with macroslip. Another model based on the Mindlin's approach fitted the loops without macroslip. Furthermore, a concept of optimizing the fixture layout by means of machine learning is presented. The intelligent combination of machine learning and finite element analysis in the design phase offers a great potential for improving the machining quality. [ABSTRACT FROM AUTHOR]

Casting process simulations are commonly used to predict and avoid defect formation. Their integration into structural optimization can enable automated structure- and process-optimized castings. Nevertheless, these simulations are time-consuming and computationally expensive. Therefore, this paper used graph theory and skeletonization techniques to extract geometric features from arbitrary 3D geometries and transferred them to machine learning-metamodels. This method can replace casting process simulation for the prediction of directional solidification in low-pressure die casting. Automated machine learning and hyperparameter optimization were used to systemize the search for well-suited neural network architectures. Two examples were used to train the metamodels, which are subsequently evaluated by a further test example, unknown to the training data and compared to the simulation results. The results showed an accuracy on unknown geometries over 60 % and thus emphasized that neural network metamodels are capable of replacing time-consuming casting process simulation for specific objectives. [ABSTRACT FROM AUTHOR]

Machine Learning (ML) offers significant potential for quality management in production with predictive analytics. Key aspects to building ML models are the selection and engineering of features from data. They allow the usage of relevant data for training ML models. Using the right features consequently improves the quality of the ML models. However, feature engineering requires knowledge of the data, data preprocessing techniques, algorithms, the domain, and use case. Hence, automatic feature engineering tools have become popular. In this paper, we investigate how domain experts and automatic tools compare for engineering features based on a time series dataset from production. [ABSTRACT FROM AUTHOR]

Industry 4.0 is accompanied by various technologies, which offer great potential for optimizing today's manufacturing of electric motors. To meet high-quality standards and to enable further improvements, methods from artificial intelligence, in particular machine learning (ML), are recently moving into focus. Although needle winding is one of the most widely used processes for winding stators with inner slots, the potential of ML has not yet been tapped. Therefore, this paper introduces a novel, data-driven approach for quality monitoring of needle winding processes using ML techniques. To acquire quality-related process data, a common needle winding machine is equipped with suitable sensors first. Based on this, a proof of concept for an ML-based regression model aiming at predicting the coil quality solely based on process data is provided. [ABSTRACT FROM AUTHOR]

In the past, research in the production domain was driven by mathematical and physical description of production technologies. Over the last years, data-driven approaches like machine learning (ML) and artificial intelligence (AI) gave the research a new direction. Often, already existing knowledge is neglected when using data-driven approaches resulting in models that do not represent the best possible results. By combining these two approaches all available knowledge is used generating the best possible model. This combination is called hybrid modeling. In this paper, hybrid ML as part of hybrid modeling is introduced and the benefits and challenges using hybrid ML for the prediction of process parameters in the production domain are demonstrated. [ABSTRACT FROM AUTHOR]

The growing use of Renewable Energy Sources (RESs) is bringing the energy efficiency of factory processes more and more into focus. The prediction of the process-specific energy demand is indispensable owing to the need to use RESs systematically. In addition, the demand to utilise cross-sector energy resources through sector coupling has also necessitated an increase in need for accurate forecasting of production processes and its energy consumption. Furthermore, the occupancy and operational times of factory machines and their respective schedules also plays a vital role in determining their energy costs. However, detecting the machine states for energy-specific needs is a challenging task due to differing machine-specific energy usage and aggregated load monitoring. Hence, this paper investigates the capabilities of clustering methods for identifying different process steps of Selective Laser Melting (SLM) based on cluster analysis of non-intrusive electrical measurements. The resulting approach enables active monitoring of SLM process steps electrically and to develop energy optimised production models thereby further reducing energy costs per production cycle. [ABSTRACT FROM AUTHOR]

In metal-cutting production, the service life of milling tools must be optimally utilized. Machine learning methods are suitable for classifying the wear of milling tools. In this paper, a convolutional neural network is presented that classifies the wear of milling tools mapped on image data with an accuracy of 94.41 %. In addition, the ability of the convolutional neural network to detect relevant areas within the image is analyzed using gradient-weighted class activation mapping. The convolutional neural network is able to detect relevant areas in the image data, even though these areas are not marked in the input data. [ABSTRACT FROM AUTHOR]

Monitoring systems may contribute increasing the availability of machine tools and detecting process deviations in time. In the past, machine learning has been used to solve a variety of monitoring problems in machining. However, boundary conditions for the assessment of the principal applicability of machine learning approaches for supervised anomaly detection in machining have not been exhaustively described in the literature. In this paper, objectives as well as deficits of literature approaches are identified and influencing factors on the monitoring quality are described. As a result, we derive boundary conditions and discuss challenges for successful implementation of machine learning based monitoring systems for supervised anomaly detection in industrial practice. [ABSTRACT FROM AUTHOR]

With the increasing amount of available and connected data sources, industrial automation applications such as condition monitoring of a production machine can be improved by considering various data. To gain insights from this data and make it useable, heterogeneous data has to be analyzed intensively. Limited machine learning approaches exist in industrial automation and manufacturing for analyzing data acquired from multiple sources. In this paper, first, a suitable concept for handling heterogeneous data from integration to analysis is presented as well as a multi-layer architecture for the concept's realization. The architecture encapsulates functionalities into the different layers and allows easy extendability and modifiability. Afterwards, a context modeling approach for managing heterogeneous data and existing approaches and algorithms for analyzing this data robustly and dynamically analyzing it are presented. [ABSTRACT FROM AUTHOR]

Cloud Manufacturing matches supplier capacities and customer demands via an algorithmic agent. The agent aims to maximize its own and the platform participant's revenue, hence the platform's gross revenue. Volatilities in supply and demand, as well as individual pricing models and willingness to pay, exacerbate the problem for deterministic solvers. In Revenue Management, comparable problems are approached with linear programming and regression models. Within this paper, we introduce a new Machine Learning (ML) based approach: Machine Learning for Cloud Manufacturing Cost Prediction (ML-ConCord). ML-ConCord aims to improve prediction quality for Revenue Management in Cloud Manufacturing by combining a recurrent neural network (RNN) with a linear programming pricing algorithm. We evaluate our approach on a self-generated dataset and compare its performance with standard regression-based approaches. Models and datasets are published on GitHub. [ABSTRACT FROM AUTHOR]

When using a force measuring unit, measurement errors occur because of deformations of the cover plate due to different force application points in the workspace. This paper deals with the question to what extent machine learning approaches are suitable to compensate such measurement errors. Two approaches, the deep neural network and multiple linear regression, are investigated. The programs created receive as input the measurement signals of a force measuring unit equipped with strain gauges and return as output estimated values for the acting force vector. Finally, the results of the two algorithms used are compared. [ABSTRACT FROM AUTHOR]

Material flow simulation is essential to modern manufacturing engineering. Physically based, discrete material flow simulation is slow, and, because it is not well suited for parallelization, it does not benefit from vertical scaling of the computer infrastructure. Other methods, like based on hyperbolic partial differential equations or Navier-Stokes based continuous simulations can be significantly faster, however switching between continuous and discrete representations is not yet solved. Machine learning methods have been used recently to speed up physically based simulations. In this paper we show, that the latest advances in machine learning solutions using graph networks can be applied to speed up industrial material flow simulations while still using discrete representations for the material flow objects. [ABSTRACT FROM AUTHOR]

Deploying machine learning (ML) models in production environments comes with challenges such as the model's integration into live production and the missing trust of process experts in new technologies. These challenges must be addressed already in phases ahead of the deployment. Therefore, this paper aims to clarify how to ensure the deployability of methods used during model development. For this purpose, criteria for measuring and evaluating deployability in manufacturing environments are defined. A subsequent analysis of existing data preprocessing methods and ML algorithms regarding deployability as well as deployment options serves to counteract deployment issues early on in an ML project. [ABSTRACT FROM AUTHOR]

The machining of Fiber Reinforced Polymers (FRP) is accompanied by specific defects such as delamination, uncut fibers, and others. Such defects are unique in their shape and size for different FRP types, used tools, and applied cutting conditions. Therefore, defect recognition and quantification remain a central challenge in the quality control of FRP components from an accuracy and time balance perspective. The study presents the implementation of Machine Learning techniques for automated recognition of the hole defects resulting from drilling Flax/PLA biocomposites using HSS drills with various cutting data. The paper discusses the effectiveness and stability of the developed solution. [ABSTRACT FROM AUTHOR]

A predictive cutting force model is critical during the end milling of Carbon Fiber Reinforced Polymer (CFRP) composites to achieve optimum fiber delamination, tool wear, and surface accuracy. The mechanistic force model showed robust prediction abilities; however, cutting constant estimation is challenging due to the involvement of multiple variables resulting in higher-order mathematical formulations. This paper presents augmenting the Feed Forward Neural Network (FFNN) with a mechanistic force model to approximate independent relationships of the chip load and edge contact length with shearing and rubbing constants. The prediction abilities of the proposed approach are compared with the traditional Fourier fitting model and experimentally measured values over diverse cutting conditions. It has been demonstrated that the proposed approach can accurately estimate cutting forces during the end milling of CFRPs. [ABSTRACT FROM AUTHOR]

Predictive maintenance is an enabler of achieving a sustainable production in 3D printing processes. The goal of this paper is to answer the question how to realize predictive maintenance for Fused Deposition Modeling (FDM) of Polyether Ether Ketone (PEEK). In a real-life use case, aging effects of the 3D printer's nozzle have a direct negative impact on product quality but are not detectable from the outside. For this purpose, a machine learning pipeline is built including the integration and preparation of sensor data as well as the training of a regression model to predict the nozzle's remaining useful lifetime. [ABSTRACT FROM AUTHOR]

Artificial Intelligence (AI) has made significant progress in supporting circular economy and reverse logistics by learning from diverse data to predict, e.g., routes or to assist workers in sorting. However, it remains an open question how AI can be integrated and trained into such operational processes, where little to no data has been collected previously. Traditionally, AI models would only be rated by their accuracy. This paper aims to introduce the concept of green incremental learning, i.e. rating AI models not only for their accuracy but to evaluate energy efficiency as well. A ramp-up of a data-driven AI system for part recognition is explored under consideration of energy efficiency. Therefore, we combine online and incremental learning, working with growing data sets to simulate a ramp-up phase. We present experiments of incremental learning on business and image data, partially supported by regular joint training steps. We start local CPU-based machine learning and prediction on business data from the first sample. Finally, we compare incremental learning to traditional batch learning and show energy-saving potential of up to 62 % without a significant drop in accuracy. [ABSTRACT FROM AUTHOR]

With the increase in the number of avenues for production of data everywhere including logistics, using this data has become an obvious next step. In this paper, we look at the major logistics problems in details as discussed in a seminal work in highlighting the problems in logistics by X. Li in 2014. Together with discussing the problems, we also look at the trends in solving these problems using machine learning. We look at the recent work done in the particular fields using machine learning (ML) and develop the correlation of the trends of using supervised, unsupervised and reinforcement learning in solving these problems. This correlation is developed using a table and a diagram showing ML techniques and the trends in the fields of the problems are discussed. COVID-19 has greatly accelerated these trends. This paper serves as a gentle introduction to ML techniques in the field of logistics for researchers who are new to the field. [ABSTRACT FROM AUTHOR]

Manufacturing evolved substantially over the years and with it the inherent complexity increased well. Currently organizations are transitioning towards smart manufacturing and implementation of industry 4.0 paradigms. Artificial intelligence and machine learning can help organizations of any size on this journey. The physics of manufacturing processes have been extensively studied and documented as relationships and equations. Machine learning algorithms rely on quality data to enable inferences and smart actions. At the intersection of these two methods lies hybrid analytics which captures the knowledge in the defined physics-based relationships and uses data-driven methods to fill the gaps that the physics-based model may have. The goal of this paper is to identify approaches and use cases of hybrid modeling wherein the term hybrid is used to indicate the combination of a physics- or first-principle-based modeling component and a data-driven modeling component. VOSViewer, Voyant Tools, and Microsoft Excel have been used to understand the relationships between various attributes associated with the articles on hybrid modeling. The papers cited the most, the organizations/regions that funded the research in this domain, the contents of recent papers, and what categories and keywords are frequently used are discussed. This work is the basis for an extended literature review to understand in detail the algorithms, use cases, and implementation issues of hybrid analytics. This bibliometric analysis uses the state of the art references and procedures mentioned in the documentation of VOSViewer and Voyant Tools. [ABSTRACT FROM AUTHOR]

Electric traction motors must meet high requirements in terms of efficiency in driving operation as well as cost-effectiveness in manufacturing. The so-called hairpin technology yields potential for both aspects. However, a central bottleneck in hairpin stator manufacturing is the contacting process, commonly performed by laser welding, where a large number of joints must be created. Since a single defective weld can lead to a failure of the whole stator, effective methods for quality assessment are needed. Building on preliminary work, this paper proposes a novel multi-view deep learning architecture for combined processing of pre- and post-process images to detect welded joints with an insufficient cross-section. [ABSTRACT FROM AUTHOR]

Machining is one of the final steps in the manufacturing value chain, where the dimensional tolerances are fine-tuned, and the functional surfaces are generated. Many factors such as the process type, cutting parameters, tool geometry and wear can influence the surface integrity (SI) in machining. Being able to predict and monitor the influence of different parameters on surface integrity provides an opportunity to produce surfaces with predetermined properties. This paper presents an overview of the recent advances in computational and artificial intelligence methods for modelling and simulation of surface integrity in machining and the future research and development trends are highlighted. [ABSTRACT FROM AUTHOR]

Enameled copper wire is a crucial material for the production of modern electrical machines. With the advent of hairpin windings, copper wire with rectangular cross-section is gaining importance. Being delivered on coil carriers, the wire is subject to initial curvatures which must be leveled out before further processing. Since common straightening systems are set manually and not controlled automatically due to missing sensor feedback, the straightening process is very susceptible to material-related deviations. After providing an in-depth problem statement, this paper gives a structured overview of related approaches and derives a basic concept for the intelligent straightening of flat enameled copper wire. [ABSTRACT FROM AUTHOR]

The efficient execution of process planning activities requires the knowledge from several distinct domains. However, in this context, a common issue is the existence of collective knowledge in the one domain with a lack of networked expertise with other domains. Motivated by this, the paper proposes an approach to support process planning by comparing the generated part design with older, validated products. The utilization of this earlier CAD part models has the potential to reduce development costs, shorten the production start-up time and improve the product quality. The new concept provides the manufacturing personnel with a method for comparing the newly designed part with a pool of validated models to identify the most similar one. Each previous CAD part model is linked with the necessary manufacturing information, so that the initial values for the manufacturing process are available without a time consuming testing phase. The method is divided in three main steps: the global similarity comparison, the segmentation of the part and the local similarity comparison. In the first step, the geometry is projected onto a sphere and then transformed to a matrix. Afterwards these matrices are compared and clustered into corresponding groups. In the following step, a Machine Learning algorithm segments the objects into specific, manufacturing relevant groups. In every cluster, the segmented geometries are again compared for similarity. The combination of the first and the second ranking results in a global similarity hierarchy for the newly designed part. In this paper, the entire procedure is shown with the example of sheet-bulk metal formed parts. This new manufacturing process particularly benefits from this method, as the amount of data is still limited and therefore little expert knowledge exists. [ABSTRACT FROM AUTHOR]

The possible role of stem cells in medical treatments can hardly be overestimated. Today they are produced – almost without exemption – with significant human involvement using adaptive protocols that take the growth behavior of the biological material into account. Automated production platforms are being developed and tested in a number of research laboratories with the main goals of improving reproducibility, as well as increasing quality and throughput. However, automated stem cell production differs from the traditional manufacturing processes in (1) the inherent diversity of the products (stem cells), (2) their varying growth rates and process times, (3) the need for their regular observation and process adaptation, and, therefore, (4) for mixed-initiative production control. A distinctive feature of the domain is the symbiotic co-existence and co-evolution of the technical, ICT and biological ingredients in production structures. A challenging way to overcome these issues is the use of biologically-inspired control algorithms. In the paper the application of reinforcement learning is proposed for this purpose. As a first step, a digital simulation of the stem cell production was performed in order to generate patterns for the training process and to test the approach. In addition to the description of the concept, the paper also presents initial research results. [ABSTRACT FROM AUTHOR]

In this paper, restaurants store management based on demand forecasting is proposed. The restaurant service industry has low productivity due to the simultaneity of service goods. In order to solve such problems, we are researching how to manage restaurant stores such as employee placement, food material ordering, etc., based on highly accurate demand forecasting by machine learning with internal data such as POS data and external data exiting in ubiquitous such as weather and events. In this paper, we discuss the forecasting results of customer order quantity and shop inventory order quantity of draft beer using forecasting method with machine learning for restaurant chain R. [ABSTRACT FROM AUTHOR]