Deep learning based simulation for automotive software development

Licentiate thesis, 2022

The automotive industry is in the midst of a new reality where software is increasingly becoming the primary tool for delivering value to customers. While this has vastly improved their product offerings, vehicle manufacturers are increasingly facing the need to continuously develop, test, and deliver functionality, while maintaining high levels of quality. One important tool for achieving this is simulation-based testing where the external operating environment of a software system is simulated, enabling incremental development with rapid test feedback. However, the traditional practice of manually specifying simulation models for complex external environments involves immense engineering effort, while remaining vulnerable to inevitable assumptions and simplifications. Exploiting the increased availability of data that captures operational environments and scenarios from the field, this work takes a deep learning approach to train models that realistically simulate external environments, significantly increasing the credibility of simulation-driven software development. 

First, focusing on simulating the input dependencies of automotive software functions, this work uses techniques of deep generative modeling to develop a framework for realistic test stimulus generation. Such models are trained self-supervised using recorded time-series field data and simulate the input environment much more credibly than manually specified models. With the credibility of stimulus generation being an important concern, an important concept of similarity as plausibility is introduced to evaluate the quality of generation during model training. Second, this work develops new techniques for sampling generative models that enable the controlled generation of test stimulus. Allowing testers to limit the range of scenarios considered for testing, the Metric-based Linear Interpolation (MLERP) sampling algorithm automatically chooses test stimuli that are verifiably similar to a user-supplied reference, and therefore measurably credible. While controllability eases the design of tests, credibility increases trust in the testing process. Third, recognizing that sampling may be an inefficient process for stimulus generation, this work develops a technique that extracts properties from actual code under test in order to automatically search for appropriate test stimuli within the specified range of test scenarios. Fourth, further addressing the question of credible stimulus generation, this work introduces techniques that examine training data for biases in sample representation. Overall, by taking a data-driven deep learning approach, techniques and tools developed in this work vastly expands the credibility of incremental automotive software development under simulated conditions.