Industrial Data Analytics refers to data analyses across diferent phases of the industrial product life cycle. The specific characteristics of available industrial data often pose challenges for common data management and data analysis methods. This paper gives an overview on the projects of the research group ICT Platform for Manufacturing at the Graduate School of Excellence advanced Manufacturing Engineering (GSaME) of the University of Stuttgart. These projects are related to Industrial Data Analytics and ofer approaches to addressing the domain-specific data characteristics. Relevant research areas are metadata management, use of domain knowledge to improve data preparation, the management of machine learning (ML) models, and approaches to meta-learning and automated machine learning (AutoML). In addition, this paper details on two specific research contributions. Firstly, it discusses a metadata model that facilitates a democratized access to data in virtual product development projects. The second contribution is an approach to exploit domain knowledge during data preparation in order to address two of the most important challenging data characteristics in industrial data: a multi-class imbalance and a data bias that is due the high variety of underlying products.
CEUR
ceur-ws.org
1. Introduction
Industrial Data Analytics refers to problems and solution
approaches to data management, data provision, and
data analytics across diferent phases of the industrial
product life cycle [
The specific characteristics of available industrial data pose challenges for common data management and data analysis methods [6, 7, 8, 9]. For instance, data may come in diverse and heterogeneous formats and be contained in isolated data silos across diferent organizational units of a company. This makes it dificult or even impossible to over, the high product diversity increases the number and complexity of patterns and correlations contained in learning (AutoML) [18, 19].
After giving an overview on related research projects
of the group ICT Platform for Manufacturing in Section 2,
this paper details on two specific contributions in
Sections 3 and 4. The first contribution is a metadata model
that connects data from heterogeneous and previously
The second major contribution is an approach to exploit
domain knowledge from a taxonomy during data
preparation in order to address two of the most important
challenging data characteristics and kinds of bias in
industrial data: a multi-class imbalance and a data bias that
is due the high variety of underlying products [
CRM … & tr sageU uppoS 2 7 3 MDA/PDA 6 CAQ
4 5 CPS MES CAP SCM
D lvee rP opem tcduo 1 n t
1 Activities in Virtual Product Development Projects
2. Research Group ICT Platform
for Manufacturing
and configurations of ML tools and software to perform,
e.g., data collection, data preprocessing, model training,
and model deployment [21]. One contribution of this
Figure 1 gives an overview on the projects and research project is a method to structure the collabaration among
topics related to Industrial Data Analytics in the research diferent stakeholdes, e.g., data scientists, IT experts,
busigroup ICT Platform for Manufacturing. Furthermore, the ness analysts, engineers or other domain experts during
ifgure assigns the projects to the phases of a typical prod- development projects for ML solutions [22]. In
addiuct life cycle, from which the projects mainly acquire and tion, AssistML is a novel concept that enhances the ML
analyze data. The first project deals with metadata mod- model management platform of the previous project by
eling in the area of virtual product development [
The major outcome of the third project is a platform their application to fault detection and fault diagnosis in to manage machine learning (ML) models, e.g., classi- a production line [15]. Hybrid approaches combine dataifcation or regression models [ 16]. Data scientists of a driven models with knowledge-based models, such as company may develop ML models for specific use cases physics-based simulation models. These diferent kinds and upload them in the platform. The ML models are of data-driven and knolwedge-driven models enhance then associated with appropriate metadata to further de- and complement each other. This is particularly useful scribe them. This metadata covers, amongst others, life in cases when pure data-driven solutions are not adecycle information of the ML model, e.g., whether it is quate due to a lack of data. A major contribution of this in the training or application phase or whether it has project is PUSION, a generic and automated framework already been retired and is thus not used anymore [20]. for decision fusion in classification ensembles that may Moreover, the metadata covers semantic information [16]. be composed of diverse data-driven or knowledge-driven This for instance includes information about the domain- models [18]. By combining the decisions of these diverse specific use case, e.g., fault detection or fault diagnosis, or classification models via decision fusion algorithms, the about the machine in a production line for which the ML overall prediction accuracy may be enhanced. model has been developed. It helps other stakeholders The focus of the next project is related to methods ifnd appropriate models for their specific use cases and for data-driven prediction of product failures during the thus efectively facilitates reusability of ML models. product usage phase [5]. Here, diferent kinds of data
The contributions of the fourth project ofer struc- set shifts may occur, i.e., changes of the statistical data tured methods to specify, configure and select whole ML distribution over time [23]. This means that both the solutions. These ML solutions constitute combinations decision boundaries of classification patterns and the ltua taa x d tnoeC teaM s e rko iiit
v W tc
A s r taaD itnena o
C ltua taa x d tnoeC teaM factory:
② Product Planning
④ Product Simulation
⑥ Prototyping ① Product Specification
③ Virtual Prototype
⑤ Simulation
Result
⑦ Physical Prototype … ⑧ Product Testing owner: Mrs. X … dimensions: 3 voxels: 30x40x50 … … … material: paper … ⑨ Test Results … … statistical distribution of these patterns may change. This or CSV-formated files. In addition, diferent data are has to be reflected via adequate approaches to detect data contained in isolated data silos across individual organiset shifts and to adapt classification models to the new zational units that are often not willing to share their data statistical distribution if necessary. with other stakeholders, even from the same company.
Finally, one project concerns data quality of text data. Altogether, this makes it nearly impossible for domain exIt introduces the QUALM concept for continuous data perts, i.e., development engineers, to acquire and explore quality measurement and improvement at diferent steps the data they need for a particular data analysis. of text analysis pipelines [12]. QUALM data quality indi- Ziegler et al. [13] thus propose a metadata model to cators quantify text characteristics, e.g., the number of describe all related CAx data and to address the aboveabbreviations or spelling mistakes, and give hints how mentioned challenges. This metadata model not only these may afect the quality of analysis results. Corre- describes data, but it ofers a connected view on data, sponding QUALM modifiers use these text characteristics metadata, and work activities in virtual product develto enhance the text quality. An example of such a modi- opment projects. Figure 2 shows an idealized example ifer is an approach to select the best-fitting training data of an instance of the metadata model. It covers several for an analysis task based on the similarity between this data containers [27, 28] (blue in the figure) that abstract training data and the input data [24]. In addition, QUALM from heterogeneous data formats and point to the unofers a hybrid method for information extraction, which derlying data sources or files via URIs. Furthermore, the exploits both structured and unstructured data sources to metadata explicitly describes the work activities (yellow) yield more relevant information from these sources [25]. that are carried out by development engineers in virtual product development projects. These work activities are connected to the data containers the activities con3. Metadata Model for Virtual sume and produce. So, the whole metadata describes full Product Development Projects and connected views of project workflows including the activities and the data. The grey elements shown in FigProduct development projects in companies are mainly ure 2 represent metadata that further describe either data virtual and digitized thanks to several computer-aided containers or work activities via contextual information. systems, e.g., for Computer-Aided Design (CAD), A major benefit of this metadata model is that product Computer-Aided Engineering (CAE), and Computer- development engineers easily understand it, because it Aided Testing (CAT) [26] These CAx systems produce a is based on the project workflows and work activities huge amount of data that ofer additional opportunities these engineers carry out in their daily work life. This for data analysis, e.g., to gain insights how to improve facilitates a democratized data access, so that product dea product design. However, several challenges hinder velopment engineers may easily find the data associated exploiting the full potential for data analysis [13]. In par- to the work activities in development projects they are ticular, diferent CAx systems store their data in various familiar with. It facilities an expert-led data exploration heterogeneous formats, e.g., proprietary 2D or 3D ge- and a subsequent data analysis to gain sophisticated inometry files, plain text, images, videos, XML documents, sights from the underlying data. The metadata model after SPH
s p u o r G t c u d o r P
Heterogeneous Segmentation according to Product Hierarchy
(SPH)
4. Approach to Exploit Domain
Knowledge for Data Preparation Hirsch et al. [11] introduce a challenging use case of data-driven identification of quality issues in complex assembly products, e.g., truck engines. The use cases constitutes a multi-class classification problem, where each class corresponds to one of 84 engine components, e.g., cylinders, fuel injectors, or turbochargers. The problem is to train a classification model that is able to identify one of these multiple classes and engine components that are the cause of a particular quality issue.
With 1050 data instances, the data set of this use cases is of rather low size and contains several kinds of noise [11]. Nevertheless, literature comprises techniques which are able to deal with these two data characteristics. However, such data-driven methods usually still show poor prediction performance, as they are not able to address two additional kinds of domain-specific data ance, i.e., the class labels occur in an imbalanced way in the data [32]. Here, many learning algorithms tend to ignore the patterns of class labels that are underrepresented. The second challenging data characteristic results from the fact that companies ofer
heterogeneous product groups with a high product variety. Here, the class patterns, i.e., decision boundaries in the feature space of a particular class usually difer across individual product the class patterns of seldom product groups that are underrepresented in data.
Hirsch et al. propose an approach to data preparation
that efectively addresses both challenges arising from a
multi-class imbalance and from heterogeneous product
groups [
The first step, Segementation according to Product Hierarchy (SPH), uses domain knowledge from a product hierarchy or product taxonomy to divide the data set into one subset for each product group. As only data of one particular product group is contained in each subset , this subset contains a significantly less number of class patterns, i.e., usually only one pattern for each remaining class in . In addition, these class patterns are more evenly distributed in each subset. So, learndistinguish all class patterns, even those that have previously been underrepresented in the whole data set
The second step, Class Partitioning according to Imbalance (CPI), further divides some of the subsets to mines the degree of class imbalance in each subset resulting from SPH using the Gini coeficient as an imbalance metric. If the value of the Gini coeficient as higher a subset than a threshold, e.g., 30 %, the subset is divided into
+ containing only data instance of majority classes and a subset − for instances of minority classes.
Here, CPI uses a quantile approach to determine the point of intersection between majority and minority classes. for classification ensembles [ 30], e.g., Random Forest [31], ing algorithms have much less problems to identify and characteristics [11]. The first one is a multi-class imbal- address multi-class imbalance. Therefore, CPI first deter
Hirsch et al. [