Combining knowledge graphs with an MBSE approach to identify and analyse potential errors in technical products

1. Introduction

Technical products are developed to meet specific requirements. For this purpose, the product must provide relevant functions in combination with other relevant properties (e.g., reliability, availability, etc.). In this context, the function is a transformation of intended (function-relevant) input flows (E_f) into intended (function-relevant) output flows (A_f) (see Figure 1). The implementation of products leads to limitations in the realization of the functions, since the product characteristics (according to the CPM model [1]) have, among other things, deviations or aging. In addition, external disturbance input variables E_n (e.g., vibrations) have an impact on the behaviour [2].

Developing products that are as robust as necessary (see Hansen's harmlessness condition [3]) is a key objective. In this context, robust means that the product must comply with the required properties of the relevant output flows (intended and unintended) within acceptable tolerances even in the presence of variations in the input variables and the internal parameters. The deviations between the actual and required properties are called errors [2, 4].
To analyse possible failures in the context of product development, the interactions between external input variables and internal state parameters must be investigated based on cause-effect relationships in the product. As a basis for the analysis, an appropriate system description based on systems theory concepts is useful.
A widely used approach for modelling complex mechatronic products is described in the so-called Model-Based Systems Engineering (MBSE) [5] using the standardised language SysML [6]. These models mainly represent the content aspects of the product to be developed through model elements (use cases, requirements, function, logical elements, etc.), their parameters and relationships, which can be used as a basis for analysis [7]. These models need to be mature before they can be used to analyse potential failures. This leads to efforts in product development. In addition, many insights into the behaviour of the product to be developed in the planned environments are often not yet available [8].
The approach presented intends to reduce the modelling effort and increase the model quality by enhancing the system models with existing knowledge from other development processes and products using knowledge graphs.
The following research question arises: How can existing knowledge about existing products and their relevant properties be used to enhance system models of the system of interest (SOI) using knowledge graphs as a basis for the analysis of possible errors?

2. Overall Approach

As a basis for the detailed explanations, the overall concept is presented first (see Figure 2). The developed approach should support engineers in analysing potential errors based on MBSE models in conjunction with a knowledge graph during the concept phase of product development. The proposed approach consists of following steps:

1. Definition of the system model of the system of interest (SOI) based on the initial state of knowledge in a project using modelling with SysML.
2. Analysis of the system model to obtain the necessary input information for the query in the knowledge graph.
3. Search for existing knowledge on disturbances or known errors in the knowledge graph and transfer this knowledge to the system model.

Step 1 consist of two steps: creating (a) black- and (b) white-box system models (see Figure 2: Right). In the black-box model, the problem space of the SOI in the planned environment is defined. The problem space includes the use cases that the SOI must provide, considering existing constraints (i.e., environmental objects). The white-box model is created based on the black-box, in which the system is decomposed into relevant sub-systems considering the interfaces to the identified environmental objects in the black-box model, forming a high-level system architecture model [9]. The black- and white-box models are the basis for the query of possible errors in the knowledge graph, since the combination of black- and white-box model describes the expectations of the system behaviour and their realization in the system (e.g. a required detection of disturbance objects (black box) is realized in the system by means of specific measurement functions and sensors).
In step 2, the black- and white-box model is used as a basis for determining the necessary inputs for the search queries. For this, the already defined model elements and their relations are traversed. In step 3, the potential errors can be searched based on the captured model information by querying (e.g., query(q) as shown in Figure 2) into the knowledge graph.
For this contribution, potential errors from already existing products were added into the knowledge graph to create the basis for the search (see Figure 2: left). This aspect is not described in detail here.
The proposed approach is evaluated using an example of a mobile robot system. For this purpose, a knowledge graph framework is coupled with the SysML modelling tool via a plugin, which is further in the development phase, to automate the proposed method. The plugin directly supports stakeholders in the SysML modelling tool (i.e., Cameo Systems Modeler) to analyse potential errors based on the defined system models.

3. Querying the Knowledge Graph based on the Captured Information from System of Interest Analysis

A knowledge graph is queried to retrieve potential errors based on the information captured from the high-level SysML model. To achieve this, an RDF knowledge graph [10] is created by considering the knowledge about potential errors. To build an RDF knowledge graph, RDF datasets are filled with triples as an example [11]. The defined RDF datasets are included with potential errors based on empirical knowledge from previous projects. These RDF datasets are stored in the knowledge graph database, which supports the visualization of the RDF knowledge graph, as shown in Figure 3.

To retrieve potential errors based on the captured input information of the SysML model, the RDF knowledge graph, that is realized by stored RDF datasets, is queried. The query is performed using the SPARQL query language [13] by accessing the SPARQL endpoint generated by the graph database. In Figure 4, the captured input information, such as the “Detect Obstacle” function is used to detect “Obstacle” with “Velocity” supports in analysing the potential errors (e.g., Velocity Too High) by querying the RDF knowledge graph that is realized by RDF datasets via the SPARQL endpoint.

4. IMPLEMENTATION DETAILS

For the coupling of the SysML tool Cameo Systems Modeler (CSM) from Dassault Systemes [14] with the knowledge graph, the programming interface of the CSM is recommended. A plugin was created in the CSM for this purpose. CSM provides a plugin development framework that supports customizing or extending the functionality of the modelling tool through its application programming interface (API). The API offers interfaces, classes, methods, etc. to interact with the main features of the CSM using the JAVA programming language. In this contribution, an initial prototype of the CSM plugin is presented, and a fully functional plugin is currently in the development phase. 

The initial prototype involves establishing a connection between the CSM modelling tool and the knowledge graph database. The developed connection retrieves potential errors within the CSM by querying the RDF knowledge graph obtained from RDF datasets. Before explaining the implementation details of initial plugin prototype, the construction of the RDF knowledge graph is discussed. The Ontotext GraphDB [15] is selected as knowledge graph database because it efficiently stores, retrieves, and manages the RDF datasets. An RDF is a data model that represents information in triples. The relationship between these triples forms a graph structure specified as an RDF knowledge graph. The implementation details comply with the proposed approach; therefore, the RDF datasets are defined based on the input information captured from the existing SysML model. Moreover, the defined datasets are loaded inside the Ontotext GraphDB repository that provides the feature to visualize the RDF knowledge graph (see Figure 6: Ontotext GraphDB part). In addition, Ontotext GraphDB provides a SPARQL endpoint (i.e., via the Ontotext GraphDB API) that acts as a means of interacting with the RDF knowledge graph realized by RDF datasets.
The implementation details of the initially developed prototype begin by exporting the required configuration files and loading them into the CSM plugin directory (see Figure 6). The predefined SPARQL query written in the initial prototype plugin is performed based on the selection of activity of the SysML model through the customized icon. The formulated SPARQL query is sent to Ontotext GraphDB (i.e., SPARQL endpoint) via an HTTP request. The Ontotext GraphDB processes the SPARQL query and retrieves the corresponding information regarding potential errors from the RDF knowledge graph by realizing the stored RDF datasets. The SPARQL query response is sent back through the HTTP request protocol in JSON serialization format. Moreover, the potential error information is represented using the CSM tool interface after parsing the JSON response (see Figure 4).