Skip to main content
Fachgebiete
Automobil + Motoren Bauwesen + Immobilien Business IT + Informatik Elektrotechnik + Elektronik Energie + Nachhaltigkeit Finance + Banking Management + Führung Marketing + Vertrieb Maschinenbau + Werkstoffe Versicherung + Risiko
DE
EN
Themenseiten
Organisationspsychologie Projektmanagement Marketing Smart Manufacturing
Bücher
Zeitschriften
Weitere Formate
Podcasts Webinare Technik Webinare Wirtschaft Kongresse Veranstaltungskalender Awards
Jetzt Einzelzugang starten
Zugang für Unternehmen
Referenzkunden
ATZ-Fachkongress

Chassis.tech plus 2024


4 Kongresse in einer Veranstaltung

Treffen Sie in München die Fachleute von Chassis, Lenkung, Bremsen sowie Reifen/Räder.
Erweiterte Suche
Anmelden
nach oben

2024 | Buch

Introduction to Multimodal and Tensor Data Analytics Kapitel lesen Erstes Kapitel lesen

Multimodal and Tensor Data Analytics for Industrial Systems Improvement

herausgegeben von: Nathan Gaw, Panos M. Pardalos, Mostafa Reisi Gahrooei

Verlag: Springer International Publishing

Buchreihe : Springer Optimization and Its Applications

Enthalten in: Springer Professional "Wirtschaft+Technik" , Springer Professional "Technik" , Springer Professional "Wirtschaft"

Einloggen, um Zugang zu erhalten
insite
SUCHEN

Über dieses Buch

This volume covers the latest methodologies for using multimodal data fusion and analytics across several applications. The curated content presents recent developments and challenges in multimodal data analytics and shines a light on a pathway toward new research developments. Chapters are composed by eminent researchers and practitioners who present their research results and ideas based on their expertise. As data collection instruments have improved in quality and quantity for many applications, there has been an unprecedented increase in the availability of data from multiple sources, known as modalities. Modalities express a large degree of heterogeneity in their form, scale, resolution, and accuracy. Determining how to optimally combine the data for prediction and characterization is becoming increasingly important. Several research studies have investigated integrating multimodality data and discussed the challenges and limitations of multimodal data fusion. This volume provides a topical overview of various methods in multimodal data fusion for industrial engineering and operations research applications, such as manufacturing and healthcare.Advancements in sensing technologies and the shift toward the Internet of Things (IoT) has transformed and will continue to transform data analytics by producing new requirements and more complex forms of data. The abundance of data creates an unprecedented opportunity to design more efficient systems and make near-optimal operational decisions. On the other hand, the structural complexity and heterogeneity of the generated data pose a significant challenge to extracting useful features and patterns for making use of the data and facilitating decision-making. Therefore, continual research is needed to develop new statistical and analytical methodologies that overcome these data challenges and turn them into opportunities.

Inhaltsverzeichnis

Frontmatter
Introduction to Multimodal and Tensor Data Analytics
Abstract
In recent times, the pervasiveness of multimodal data, particularly within the scope of industrial engineering and operations research, has grown exponentially. A myriad of research has focused on integrating such data using various innovative techniques, highlighting numerous facets of multimodal data fusion, and unveiling a series of open challenges still awaiting solutions. This book sheds light on various methodologies centered on the fusion of multimodal data, particularly emphasizing the role of tensor-based data analytics. It offers a comprehensive perspective on real-world applications (e.g., manufacturing, healthcare, and renewable energy) while presenting several unique methodological domains, including functional and tensor data analysis, spatiotemporal data analytics, deep learning, federated/distributed learning, and integration of domain knowledge. The capabilities and distinguishing traits of these methods are also summarized in this introductory chapter. This section concludes with an outline that highlights the main contributions of this work and a discussion of the existing challenges and promising research avenues in the realm of tensor data analytics and multimodal data fusion.
Nathan Gaw, Mostafa Reisi Gahrooei, Panos M. Pardalos

Functional Methods for Multimodal Data

Frontmatter
Functional Methods for Multimodal Data Analysis
Abstract
Functional data analysis (FDA) encompasses a variety of statistical methodologies used to handle functional data, which may include various data modes such as time series data, spatial data, and imaging data. FDA addresses key challenges in multimodal data analysis, for instance, by summarizing, aligning, and fusing multiple data modes. In this chapter, we will discuss what are functional data and FDA, why FDA is particularly useful for multimodal data fusion, and how it can be applied to analyze multimodal datasets.
Minhee Kim
Advanced Data Analytical Techniques for Profile Monitoring
Abstract
Nowadays advanced sensing technology enables high-resolution in-process data collection during manufacturing, known as profiles or functional data. These data facilitate in-process monitoring and anomaly detection, which have been extensively studied in recent years. Yet three main challenges are the most essential: (i) how to model complex correlation structures of high-dimensional profiles, i.e., cluster-correlated or sparse-correlated profiles, (ii) how to efficiently detect changes before the profile is complete, and (iii) how to characterize the between-stage correlation of multi-stage profiles. To address these three challenges, we accordingly develop three techniques for high-dimensional profile monitoring, in-profile monitoring, and multi-stage profile monitoring.
Peiyao Liu, Chen Zhang
Statistical Process Monitoring Methods Based on Functional Data Analysis
Christian Capezza, Fabio Centofanti, Antonio Lepore, Alessandra Menafoglio, Biagio Palumbo, Simone Vantini

Tensor Analytics Methods for Multimodal Data

Frontmatter
Tensor and Multimodal Data Analysis
Abstract
In this book chapter, we provide a selective review of recent advances in tensor analysis and tensor modeling in statistics and machine learning. We then provide examples in health data science applications.
Jing Zeng, Xin Zhang
Tensor Data Analytics in Advanced Manufacturing Processes
Abstract
The emergence of edge computing, coupled with the growth of the Industrial Internet of Things (IIoT), along with sensors and intelligent/smart technologies, has opened up significant possibilities for the progression of advanced manufacturing. Together with data science and artificial intelligence, manufacturing data analytics are transforming manufacturing from limited factory floor automation to fully autonomous and interconnected systems. These data analytics methods are mainly based on vectors; however, real-world manufacturing data are presented in the format of high-order tensors. Accordingly, tensor data analytics has become a fast-growing area for advanced manufacturing. In this chapter, two robust tensor decomposition methods, motivated by specific engineering problems, are introduced for process monitoring in metal additive manufacturing.
Bo Shen

Spatio-temporal Analytics Methods for Multimodal Data

Frontmatter
Spatiotemporal Data Analysis: A Review of Techniques, Applications, and Emerging Challenges
Abstract
In recent years, spatiotemporal data has continued to proliferate with the development of data collecting technologies such as the Global Positioning System (GPS), the Internet of Things (IoT), advanced sensors, cameras, loop detectors, and various mobile applications, including social media. Efficient and effective analysis of spatiotemporal data can help extract crucial information in diversified areas such as transportation, climate and weather, the environment, human mobility, public safety, neuroscience, and epidemiology. However, with both spatial and temporal attributes, spatiotemporal data is more complex in nature, making it unique from other types of data. Consequently, additional challenges arise when working with this special data type. Nevertheless, in this era of Artificial Intelligence (AI), researchers have been relentlessly working on developing improved methods that are successful in solving various problems that require unveiling spatiotemporal patterns in the data. In this chapter, we have attempted to provide a comprehensive discussion on spatiotemporal data. We explore both traditional machine learning techniques and the currently preferred deep learning methods that are well-suited for specific problems associated with distinct types, instances, and formats of spatiotemporal data. In addition, we explore various domains where spatiotemporal data is regularly collected, stored, and analyzed. Besides, we also present a case study related to spatiotemporal track association of marine vessels using deep learning algorithms. Finally, we conclude the chapter by identifying the existing challenges associated with spatiotemporal data analysis and providing the direction to tackle these challenges in future research.
Imtiaz Ahmed, Ahmed Shoyeb Raihan
Offshore Wind Energy Prediction Using Machine Learning with Multi-Resolution Inputs
Abstract
The ever-increasing scale and penetration of offshore wind energy in modern day electricity systems is continually raising the need for wind resource and generation forecasts that are of higher quality and finer resolution, both spatially and temporally. In their quest for high-quality forecasts, a forecaster is often faced with the challenge of how to effectively make use of the wealth of heterogeneous data inputs at their disposal (e.g., on-site and off-site sensory data, multi-resolution weather predictions, and physics-based information), each characterized by different levels of accuracy, resolution, and/or fidelity. If utilized wisely, such multi-source information can collectively provide the forecaster with complementary “world views” of the local wind conditions at the wind farm site, ultimately enhancing the effectiveness of their forecasting approach. Machine learning (ML) presents a powerful approach to undertake this complex data fusion task, yet valid concerns about its “black-box-ness” and indifference to the underlying physics of wind field formation and propagation are often raised. This chapter briefly reviews the main lines of research on this front and then presents an ML-based approach to effectively integrate multi-resolution physics-based and data-driven information in order to make accurate short-term wind speed and power forecasts for the offshore wind energy areas in the US NY/NJ Bight where several Gigawatt-scale wind farms are under development.
Feng Ye, Travis Miles, Ahmed Aziz Ezzat
Sparse Decomposition Methods for Spatio-Temporal Anomaly Detection
Abstract
Anomaly detection constitutes a critical field of research, concerned with the identification of rare, atypical, or unexpected patterns within a dataset. Within the existing literature, the majority of anomaly detection techniques lack the capability to localize the anomalies. Recently, techniques such as sparse anomaly decomposition methods possess the distinctive ability to not only detect but also pinpoint the location of the anomalies concurrently. In the subsequent sections of this chapter, an exhaustive review of existing anomaly decomposition techniques will be conducted, with a particular emphasis on the smooth sparse decomposition method. Following this, several contemporary extensions to sparse decomposition methods will be explored, resulting in a discussion on the prospective directions for future research in this domain.
Hao Yan, Ziyue Li, Xinyu Zhao, Jiuyun Hu

Deep Learning Methods forMultimodalData

Frontmatter
Multimodal Deep Learning
Abstract
Multimodal deep learning has gained significant attention and shown great promise in various domains, including medical, manufacturing, Internet of Things (IoT), remote sensing, and urban big data. This chapter provides an overview of neural network-based fusion techniques in multimodal deep learning. The advantages of deep learning over conventional shallow learning methods are discussed, highlighting its ability to learn both inter- and intra-modality representations with minimal preprocessing and implicit dimensionality reduction. The chapter explores different fusion methods, including early fusion, late fusion, and intermediate fusion, and discusses their capabilities and limitations. It also examines various objectives used in late fusion, such as reconstruction error, correlation-based objectives, and semantic alignment. The challenge of avoiding negative transfer in multimodal learning is addressed, and regularization objectives and training approaches are explored. Overall, this chapter serves as a comprehensive guide to multimodal deep learning and its fusion techniques, offering insights into their applications and potential for future research.
Amirreza Shaban, Safoora Yousefi
Multimodal Deep Learning for Manufacturing Systems: Recent Progress and Future Trends
Abstract
The development of sensing technology provides large amounts and various types of data (e.g., profile, image, point cloud) to describe each stage of a manufacturing process. Deep learning methods have the advantages of efficiently and effectively processing and fusing large-scale datasets and demonstrating outstanding performance in different tasks such as process monitoring and diagnosis. However, multimodal monitoring data raise new challenges to apply existing deep learning methods to solve manufacturing tasks: (1) features across modalities contain complementary yet redundant information (inter-modal data fusion); (2) single modal data can also contain information from different viewpoints (intra-modal data fusion);(3) besides the fusion among data, domain knowledge in advanced manufacturing should also be actively fused into the feature extraction (domain knowledge fusion). This chapter provides three examples of cutting-edge multimodal deep learning methods focusing on inter-modal, intra-modal, and domain knowledge fusion, respectively. From the application aspect, they are designed to solve online process monitoring, product quality inspection, and manufacturing process design, which cover both the forward prediction task and backward optimization task in manufacturing systems. Several prospective directions in multimodal deep learning for advanced manufacturing are further discussed.
Yinan Wang, Xiaowei Yue

Integration of Domain Knowledge andMultimodal Data

Frontmatter
Synergy of Engineering and Statistics: Multimodal Data Fusion for Quality Improvement
Abstract
This chapter outlines the synergies achieved through the fusion of engineering and statistical approaches for quality improvement. It emphasizes the integration of data science and system theory, leveraging in-process sensing data for comprehensive process monitoring, diagnosis, and control. Multimodal data fusion is a key strategy for quality improvement, leading to root cause diagnosis, automatic compensation, and defect prevention. This approach goes beyond traditional aspects, such as change detection, off-line adjustment, and defect inspection. The chapter provides a concise overview of multimodal data fusion, highlights its recent developments and applications in data fusion for structured and unstructured high-dimensional data, and outlines challenges and opportunities in contemporary data-rich systems. Additionally, it explores future research directions, with a specific emphasis on harnessing emerging machine learning tools to enhance quality in systems with rich sensing data.
Jianjun Shi, Michael Biehler, Shancong Mou
Manufacturing Data Fusion: A Case Study with Steel Rolling Processes
Abstract
Production systems typically generate massive sensing data. Data fusion methods are required to transform these sensing data into valuable knowledge for process and quality improvement. This chapter provides a summary of a series of studies motivated by steel rolling processes, which addresses several aspects, including estimating the effects of process operations, predictive modeling, and unsupervised event identifications.
Andi Wang
AI-Enhanced Fault Detection Using Multi-Structured Data in Semiconductor Manufacturing
Abstract
The semiconductor industry is growing rapidly due to its key drivers, an increased chip demand for newly emerging technologies, as well as the existing ubiquity in industry and consumer goods. The equipment necessary for the manufacturing process is at the same time extremely expensive, and production processes are highly complex. To stay competitive and prevent yield loss, manufacturers are permanently trying to optimize current fault diagnostic and classification processes based on physical sensors within the production process. To enhance current approaches, fault detection must not be limited to structured sensor data; it should also include multi-structured data sources. This chapter outlines how current fault detection processes in semiconductor manufacturing can be improved by not only analyzing structured sensor data but also by including unstructured textual data, leading to an increased uptime. A multi-step algorithm is proposed, able to improve fault detection based on extracted problem statements and solutions from historical maintenance reports for an occurred failure. The performance of the introduced approach is evaluated in a simulation-based use case in the semiconductor industry, leading to an increase of 0.5% in uptime and an improvement of 12.03% in the mean time between failure, resulting in an improved overall equipment efficiency of 2.1%.
Linus Kohl, Theresa Madreiter, Fazel Ansari

Federated and Distributed Analytics Methods for Multimodal Data

Frontmatter
A Survey of Advances in Multimodal Federated Learning with Applications
Abstract
Data privacy has long been an item of emphasis for personal data. This is especially true for healthcare data, which is often multimodal (i.e., it utilizes in some fashion multiple data streams from multiple sources). In an effort to enhance the knowledge-base of privacy-preserving techniques with respect to multimodal data, we provide a survey of multimodal federated learning (MMFL). Our paper includes a thorough introduction to federated learning as well as a discussion on applications of multimodal federated learning to disease classification, autonomous driving, and human activity recognition, among others. Additionally, we describe various methodological advances in MMFL, a subset of which include extensions to supervised learning, personalization, generative models, data reduction, and feature selection. As a proof-of-concept for MMFL, we also include a novel application of federated learning to a series of physiological signals collected during simulated flights, known as the CogPilot dataset.
Gregory Barry, Elif Konyar, Brandon Harvill, Chancellor Johnstone

Bayesian Analytics Methods for Data withMultimodal Distributions

Frontmatter
Bayesian Multimodal Data Analytics: AnIntroduction
Abstract
Bayesian methods for multimodal data have attracted the interest of researchers and practitioners in a variety of real-world applications. Indeed, Bayesian statistics provides an effective framework to deal with mixtures of unimodal distributions, allowing one to incorporate prior information when available and to model posterior distributions in distinct modes. This introductory chapter presents a brief overview of the Bayesian perspective in the field of multimodal data, as well as a brief overview of salient applications. This chapter additionally offers the reader an introduction to two subsequent studies, wherein Bayesian modeling methods are presented for addressing multimodal data in the context of risk analysis and gestural human–machine interaction problems, respectively.
Marco Luigi Giuseppe Grasso, Panagiotis Tsiamyrtzis
Bayesian Approach to Multimodal Data in Human Factors Engineering
Abstract
Human behavior is complex, especially when engaging with large or ill-defined systems, and traditional statistical approaches typically cannot directly analyze and interpret human behavioral data. There are several statistical techniques that could be applied to modeling human behavior; however, because of the parallels between Bayesian thinking and how humans interact with complex systems, Bayesian analytical methods may be the most appropriate for modeling and predicting human behavior. This chapter reviews human factors research and experimentation, Bayesian thinking, and demonstrates an example of how to perform a Bayesian statistical analysis of human behavioral data in a human–computer interaction task. The data used in the numerical example comes from an experimental study on 3D gestural human–computer interaction for anesthetic tasks [1]. Anesthesia providers were asked to perform hand gestures to interact with a computer system and perform specific anesthetic tasks, and a Bayesian statistical approach was used to predict intuitive gesture choice based on system factors (e.g., contextual task) and individual factors (e.g., expertise). The data of interest for the Bayesian modeling is the intuitive gesture choice made by the participants which is categorical with many different levels. Thus, the numerical example outlined in this chapter analyzes data that is distributed multinomially.
Katherina A. Jurewicz, David M. Neyens
Bayesian Multimodal Models for Risk Analyses of Low-Probability High-Consequence Events
Abstract
This paper reviews a set of Bayesian model updating methodologies for quantification of uncertainty in multimodal models for estimating failure probabilities in rare hazard events. Specifically, a two-stage Bayesian regression model is proposed to fuse an analytical capacity model with experimentally observed capacity data to predict failure probability of residential building roof systems under severe wind loading. The ultimate goals are to construct fragility models accounting for uncertainties due to model inadequacy (epistemic uncertainty) and lack of experimental data (aleatory uncertainty) in estimating failure (exceedance) probabilities and number of damaged buildings in building portfolios. The proposed approach is illustrated on a case study involving a sample residential building portfolio under scenario hurricanes to compare the exceedance probability and aggregate expected loss to determine the most cost-effective wind mitigation options.
Arda Vanli
Metadaten
Titel
Multimodal and Tensor Data Analytics for Industrial Systems Improvement
herausgegeben von
Nathan Gaw
Panos M. Pardalos
Mostafa Reisi Gahrooei
Copyright-Jahr
2024
Electronic ISBN
978-3-031-53092-0
Print ISBN
978-3-031-53091-3
DOI
https://doi.org/10.1007/978-3-031-53092-0

Premium Partner

    Bildnachweise