Next Generation Automotive Hardware Design using Machine Learning
The Automotive Electronics industry is going through a period of rapid growth driven primarily by 4 different trends:
(a) Electrification of the powertrain: a worldwide initiative towards environment-friendly transportation
(b) Safety: Advanced Driver Assistance Systems (ADAS)
(c) Connectivity: infotainment, V2X communication, autonomous driving
(d) Shared mobility: efficient fleet management catering to changes in consumer behavior
The realization of such systems leads to an integration of an ever-increasing array of Electronic Control Units (ECUs) located in close proximity to each other, such as powertrains, wire harnesses, infotainment systems, sensors, processors, displays, etc.
Furthermore, ECUs don’t always work in unison with each other; sometimes, they interfere in each other’s performance.
These challenges are exacerbated for Electric Vehicles (EVs) where, for example, the induction motor noise may interfere with internal electronics.
Today, the automotive industry solves this problem with strict compliance specifications. The ECU designers build samples or prototypes and perform laboratory measurements that emulate vehicle conditions to get certified.
Laboratory slots are expensive, in the order of €370/hour, and the availability of slots is a major concern. Furthermore, a laboratory measurement will only give you the end result — if the ECU fails the test, the designer has little clue as to the root cause of the problem or the fix.
This results in several iterations of prototypes, which is a setback for time-to-market. Late stage fixes mostly rely on adding extra components, e.g. noise filtering or shielding components, which adds to the bill-of-materials.
The “holy grail” lies in system-level analysis and design tools that can capture complex multi-physics effects.
Imagine a future, where an automotive design engineer can make a design change in his module on his desktop, and in real-time visualize the system-level effects including functionality, electromagnetics, thermal, and power trends.
The Electronic Design Automation (EDA) industry has historically provided a tremendous boost to semiconductor design and verification, primarily by relying on the physical equations that govern its performance.
The challenge now is to scale this to the system level by modeling the automotive in its entirety, preserving the accuracy of each of its functional modules.
Traditional methods would require an impractical amount of time and computer memory to address full-systems. Therein lies a possibility to augment the traditional physical science-based approaches with data science or learning-based methods to address the system-level.
Machine-learning can be applied to system-level analysis in various stages.
Artificial Neural Networks (ANNs) can be applied toward cost-effective full-factorial, design space exploration. After the initial training period, the results can be obtained in real-time and will be orders of magnitude faster than traditional SPICE-based simulations.
Similarly, for Electromagnetic Analysis, there lies an opportunity to augment the information-rich invariant subspace, generated during the matrix-solution of base design, to rapidly accelerate the solution of subsequent design variants by 10x.
Machine learning can also be applied to diagnosis, which involves identifying the faulty module or component causing an undesired effect in the system-level performance. Unlike analysis which usually has a physical guiding equation, diagnosis in most cases requires inference from an interplay of different effects.
Support Vector Machines may be used to train a network, using data from different analysis or measurement steps to identify culprits in a failure condition.
Simyog Technology is working on these exact problems, with an innovative product suite focused on applying machine learning to bridge the gap in system-level tools.
By taking design files as input, we can predict electromagnetic interference levels at the module and full-vehicle levels, similar to the laboratory experience.
Furthermore, the analysis can be performed even before manufacturing a single prototype and can be used to root-cause design issues and often solve them with simple re-design without adding to the bill-of-materials.
The end result is a better, cheaper automobile delivered faster to market.
This article is a guest post by Dipanjan Gope, CEO of Simyog Technology / Assistant Professor, IISc, a company integrating physical science & data science to enable cost-effective & performance rich automotive electronic design through early stage failure detection, saving bill-of-materials & reducing time-to-market.
