Data analytics is vital in manufacturing for extracting insights from production data and optimising production processes. Semantic technologies including knowledge graphs (KG) proved to be beneficial for addressing challenges of transparency and explainability of analytics by ofering standardised means to describe manufacturing domains, data, analytical tasks and solutions. In this work we discuss executable KGs for industrial analytics; they can be “translated” (i.e. transformed) to executable data pipelines in a reusable and modularised fashion. In particular, we discuss how to capture analytical solutions in the form of data pipelines as KGs, and how to translate such KGs to executable data pipelines. The poster presents our framework, implementation, and preliminary industrial evaluation.
Data analytics is vital in manufacturing for extracting insights from production data and
optimising production processes. Semantic technologies including knowledge graphs (KG)
proved to be beneficial for challenges of transparency and explainability [
In particular, KGs allow to represent executable data analytical pipelines with standardized
and formal description to represent the steps in the data pipeline [
In this work we study how KGs can facilitate analytics by addressing two issues: how
to encode data pipelines as KGs – we refer to such KGs as executable KGs – and how to
automatically compute executable pipelines from executable KGs. Here the latter should ensure
the correctness of the analytical methods; the desired order of tasks within executable pipelines;
the correct number of input and output and corresponding designed dimensions. Moreover, such
computation should support typical executable pipelines for visual and statistical analytics [
In this poster paper we exemplify our solution within the domain of quality monitoring in
automatic manufacturing. In particular, in this poster we present the following. We present our
executable Knowledge Graph framework, its verification, translation and execution methods;
we discuss implementation of our solution, and present its preliminary evaluation with Bosch
manufacturing data. The experimental results show that our proposed approach is promising
in coverage and scalability aspects. This poster accompanies our In-Use track paper accepted at
ISWC’22 [
Executable Knowledge Graph Framework. We propose framework for executable KG
(ExeKG) that represents ML solutions [
We first define data, methods and tasks in this framework.
or items of information, it can be in forms such as numerals, diagrams or strings organised in diferent structures, typically relational tables or RDF database, etc. A
form of language-dependent script (such as in C++ or Python). A method takes some data which fulfils certain if ℱ( ) =
Constraints ℱ as input and can output specific data. Formally, = ℱ ( ), . A Task is the process of invoking a method by feeding it with some data that meets certain Constraints, and by doing so to obtain some other data. Formally, ⟨ , ℱ ⟩ = ℱ ( ) = , if ℱ( ) = . We call each single as Atomic task.
Some tasks have methods which are unified, while other more complex tasks can not solved by invoking a single integrated method while can be unfolded into a sequence of tasks where each task is a part of the complex one. We refer to the complex tasks as pipelines . Formally, a pipeline with input data to get , expressed as ⟨ , ℱ ⟩ = can be unfolded in the sequence { 1, 2, ..., }, where: 1⟨ 1, ℱ1⟩ = 1
, 1 ⊂ , ℱ1( 1) = ; ... ⟨ , ℱ ⟩ = , ⊂ ⋃ ̇ ∈ {1,...−1 } ∪ , ℱ ( ) =
⟶ ∈ ⋃̇ ∈ {1,..., } , ℱ = ⋂̇ ∈ {1,..., } ℱ ( ).
(1) Translation and Execution. The Table 1: Task complexity, categories and coverage. translation can be discussed with KG Type Structure Avg. #Atomic Tasks Coverage two structures of executable KGs: Visu- Linear 5 to 10 100% 1) Sequential: here each exe- alKG Multilinear 5 to 10 85% cutable KG is in the form of a Pipeline, Linear 1 to 5 100% wofhsiecqhuceonntisaisltsstroufcatusreersiecsononfeTcatsekds StatsKG MInPutaeltrgialrilanleteeladr 11500tttooo122000 899005%%% wlaittihonhoasfNaenxteTxaescku.taTbhluesK, Gthien vtroakness- MLKG ParaIlnletlegInratetegdrated MMoorree tthhaann 2200 8800%% the Python function scripts with the inputs/outputs and parameters given by DataEntity and datatype properties of KGs, according to the order defined by hasNextTask.
2) Parallel: In the case of merging two parallel structures, the translator will search preceding dependency with hasNextTask, until no preceding Task is found.
Implementation. We implemented a system for executable KG translation [
Transparency and Coverage Evaluation. Theoretical discussion: We formulate visual, statistical and ML analytics in the form of Eq. 1-3. These three forms provide a way to describe analytics tasks in a general and straightforward way, which eases the understanding of the tasks, since the users only need to understand the description once and then they can a b 9.7 31.449.457.682.1 understand similar data analytical pipelines described in this way. Thus, our approach is a step towards more transparent way that cover cases of data analytical pipelines described with Eq. 1-3. Empirical evaluation: We organised extensive workshops with the ML and non-ML experts. After discussion, we categorised most tasks of visual, statistical and ML analytics encountered in our project in groups (see Table 1), and give the coverage percentage according to our empirical cases. Observe, for all of the cases the coverage is above 80%, for some of the cases even above 90%. Besides, the users also gave their subjective evaluation on transparency with questionnaires, where they answered questions such as “I found the Executable KGs make data analytics easier to understand ” and gave scores ranging 1-5 (ranging for disagree to agree). The average score was 4.28 ± 0.47 (mean ± standardeviation) which shows good transparency. Scalability Evaluation. We evaluate the scalability of our approach by the running time of translation and execution of executable KGs with diferent complexity type (Fig. 2b). Data Description. To have controllable scope, we tested these executable KGs on a sample welding production dataset collected from a German factory. The dataset is in relational tables form after integration, containing 4585 welding operation records, 2 welding programs, performed by 1 welding machine and deals with 2 types of car bodies.
Results and Discussion. Fig. 2b demonstrates that our system scales well since it takes limited time to translate executable KGs to scripts and execute scripts. On most right hand side, we see that the translation and execution of most complex executable KGs, namely parallel integrated executable KG, only takes 6 minutes for 150 KGs, on the given data, which shows good scalability.
Conclusion and Outlook. In this poster we present our ongoing research of representing data analytical pipelines in KGs and transformation (also called “translation”) of such KGs in executable analytical pipelines. We discussed framework, verification, translation and execution with our scope of welding monitoring with a Bosch case and evaluated our approach with real industrial data and users from Bosch case, which shows promising results. In the future, we plan to generalise our approach to more cases and to host the system regularly on the Bosch environment and constantly collect more user feed-backs. We will also study more technical details such as expressivity and limitations in theory and practice, and compare with other similar work such as Yahoo! Pipes, DAGs in Spark, Tez, the PROV-O ontology.
The work was partially supported by H2020 projects Dome 4.0 (Grant Agreement No. 953163), OntoCommons (Grant Agreement No. 958371), DataCloud (Grant Agreement No. 101016835) and the SIRIUS Centre, Norwegian Research Council project number 237898.