Design Method for Integrating Trained Neural Nets with UML

2024-01-2013

04/09/2024

Features
Event
WCX SAE World Congress Experience
Authors Abstract
Content
Model-based developments have been introduced to reduce the development time for vehicle systems. Various model-based tools, including MATLAB and Simulink, have been introduced, and each vehicle component uses different tools to model assets. This makes the system complex and reduces the simulation efficiency because of the need for interfaces or converters when reusing model assets and combining parts. However, machine learning, in which neural nets are pretrained to make inferences in real time, is being applied to automatic driving and applications such as object recognition. This study developed a system in which the inputs and outputs assigned to a model were trained using neural nets, and the trained neural nets were combined with UML: Unified Modeling Language. A previous UML integration proposal integrated C/C++ code automatically generated from the models. Therefore, the previous proposal made limited use of modeling tools with automatic code generation capabilities. The learned network can be easily imported by adding input/output capabilities to the standard UML. While UML enables humans to understand the entire development system and prevent design errors, it becomes more complex when details are expressed. In general, UML has been used for overall design and has never been integrated with a detailed design. However, in this system, a detailed design is created by the designer using an arbitrary modeling tool and is integrated electronically, thereby reducing design errors. Furthermore, the use of pretrained neural nets reduces the simulation time and improves the development efficiency. There is also a proposal to easily switch between models created using UML and deep learning neural networks.
Meta TagsDetails
DOI
https://doi.org/10.4271/2024-01-2013
Pages
10
Citation
Arai, M., "Design Method for Integrating Trained Neural Nets with UML," SAE Technical Paper 2024-01-2013, 2024, https://doi.org/10.4271/2024-01-2013.
Additional Details
Publisher
Published
Apr 09
Product Code
2024-01-2013
Content Type
Technical Paper
Language
English