This paper describes the goals, (partial) results and lessons learned of the ECoVaDeVa project, a Flemish project that groups academic and industrial partners around the e cient, model-based development of software controller variants for Cyber-Physical Systems (CPSs). ECoVaDeVa's high-level goal is to apply Product Line Engineering (PLE) techniques to CPS controller design in all phases of the development lifecycle (design, simulation, testing, deployment) as an extension of existing software product line techniques. While PLE is well researched in software development, it is not clear whether these results apply to CPS controller design. The added complexity stems from the heterogeneity of models representing the system, involving plant, controller and environment, software and hardware, and virtual test benches (model-in-theloop, hardware-in-the-loop, etc.). The envisioned result of the project is a set of tools, techniques, and guidelines for the e cient management of CPS controller product variants. The techniques developed during the project are demonstrated on a common use case: a windshield wiper.
2.1
ECoVaDeVa is a Interdisciplinary Collaborative Research (ICON) project. An ICON project involves a strategic research center (including academic partners) and industrial partners. The goal is to transfer basic research to the industrial partners. In the project, a common research challenge is identi ed for all of the industrial partners. In the strategic basic research part, generic research questions are formulated and addressed to improve upon the state-of-the-art by the strategic research center. In the applied research part, the results of this research are then translated to the speci c case of each of the industrial partners in separate work packages. This requires a close collaboration between academic and industry partners. In this paper, we focus on the results of the strategic basic research, as applied research results are con dential. 2.2
In a previous research project (VARIES4), Flanders Make (formerly FMTC) and Dana (formerly Spicer) investigated variability tools with respect to their applicability in real-life mechatronic applications beyond the classical toy examples. From that project, a number of research challenges were identi ed.
The main goal of this project is to increase the e ciency of the development and validation process of Cyber-Physical System (CPS) controller software variants. This e ciency gain is achieved by bridging the gap from business software Product Line Engineering (PLE) methods and tool prototypes only demonstrated on toy problems to industrial-scale CPS software development and validation processes. More speci cally, the project aims to:
Create a methodology and provide tool support for building central variability models for CPSs with variability in hardware and software architectures and automatically generating a con guration tool from these models; Provide insights to companies in the various possibilities to implement variability in behaviour simulation languages such as The MathWorks Simulink and SISW Imagine.Lab by o ering them a decision tree for selecting the most appropriate variability mechanisms; Create a methodology and a toolbox that companies can use to automate their speci c build and validation process of CPS software variants, including the generation of Model-in-the-Loop and Hardware-in-the-Loop tests; Create a consistency tool, that allows for early detection of inconsistencies between central variability models and controller and plant simulation tools if the product line evolves due to change requests. 2.3
cepted PLE approach [
In order to validate and communicate new approaches within the basic research of the project, we established a windshield wiper controller use case. The case consists of CVMs and con guration tools, family models in Simulink, C++, MagicDraw, etc., and a fully automatic way to generate variants, covering all aspects of Figure 1. In the applied research of the project, the approaches that are investigated in the basic research are validated by the industrial partners in industry-scale environments. 3
Selected results of the strategic basic research are explained in this section. During the project, we made use of the commercial tool pure::variants5 to model the CVM and its relationship with modelling languages (RQ1 and RQ2).
Variability, Binding Times and Variant Generation. Approaches exist
for implementing variability by means of variation points in modelling languages,
but often, these are partial solutions. In the context of RQ2 and RQ3, it was
investigated how existing constructs can be used to express variation points in
relevant family modelling languages (Simulink, Amesim, SysML, etc.), and how
they can be linked to the CVM. Di erent types of variation points (optional,
alternative, multiple instances, etc.), di erent binding times (i.e., the moment in
the design work ow a variant choice is applied, for example, at compile-time),
and di erent levels of granularity (depending on the family modelling languages,
this can be topological, connection, element property, etc.) are investigated. This
includes the support for generating variants. A gap analysis for each used
family modelling language discovered missing constructs. A notable example is the
lack of support for model-time variability in Simulink, where Simulink model
variants are automatically generated from a Simulink family model. A remedy
was achieved by using rule-based model transformation for Simulink [
Consistency. In current approaches, it may be possible that variants are
generated from a product family, that result in errors when building/simulating
them. This is caused by a mistake in the family model, but such errors are
notoriously hard to nd in a product family, because of the interdependencies of
variation points within the family model. This hinders the short and predictable
product delivery times, aimed for by automatic generation of variants. In the
context of RQ4 and based on an approach for UML [
Plant Variability. For plant modelling, acausal modelling languages provide
more natural abstractions, as the physical world is acausal by nature. It is then
appropriate to model the variation points, if multiple product variants exist, in
such an acausal language as well. In the speci c case of Simscape for example,
variability mechanisms of Simulink [
Architectural Variability. In order to deal with controller software on
an industrial scale, an architectural overview needs to be maintained. Partial
solutions exist for e.g., the automotive sector, for which AUTOSAR has support
for variability [
Dissemination. We have built a physical demo of the windshield wiper use case to attract attention to the problem of variability management in CPS design. It has been exposed at the Flanders Make Symposium, the Hannover Messe and will be exposed at the MATLAB Expo 2019 Benelux. 4
Cyber-physical controller design typically follows a (derivative of a) V-model
process, in order to deal with di erent levels (e.g., concept, system, component,
implementation). Each layer comes with di erent models and tools. Some CVM
features are associated to models at the concept level (e.g., use an automatic
or manual transmission), whereas other features are only relevant further
downstream (e.g., which clutch control algorithm will be used). The clutch algorithm
feature is not yet available, and it may be unknown during concept design that
this involves a variation point. Consequently, the CVM should be split up
accordingly so that the right stakeholders have access to the right (partial) CVM
at the right time in the development process. Additionally, variation points need
to be modelled in the right model, and need to be bound at the desired moment
(i.e., binding time) in the tool chain, conforming the layer in which they are
dened. Another layer of features depends on the di erent tasks performed during
the development process: a model from which software is generated may need to
change if it is used for MiL testing. Approaches for multi-view variability exist
(e.g., [
In the ECoVaDeVa project, academic and industrial researchers have investigated the applicability of variability techniques in CPS development. Notably, speci c answers were provided for dealing with variability in a multi-discipline setting, consistency at family level, variability in plant modelling and variability for architectural models. Additionally, research challenges were identi ed related to managing variability in layered CPS design processes based on the Vmodel. The project is aligned with the core subject of STAF, as it aims to de ne methods, techniques and tools that deal with variability through model-driven engineering and model transformation.