No, Title
|
SCD 4.2: AI accelerated powertrain control
|
|||
Leader
|
BUT
|
|||
Contributing Partners
|
BUT, Infineon
|
|||
Description
|
BUT used an evaluation board provided by Infineon with the AURIX 3G microcontroller, which contains a Parallel Processing Unit (PPU), to demonstrate its capabilities in accelerating the computation of complex algorithms related to powertrain control and diagnostics. Two directions were explored.
The first direction focused on the acceleration of Finite Control Set Nonlinear Model Predictive Control (FCS NMPC) algorithms. These algorithms are known to be computationally intensive, and their complexity grows exponentially with the increasing number of prediction steps.
The second direction involved the use of AI in diagnostics, specifically the implementation of a convolutional autoencoder for anomaly detection in powertrain operation.
|
|||
Deployment/utilization
|
A simplified variant of the FCS NMPC control algorithm (one-step-ahead prediction shown in Figure 1) for controlling a three-level inverter was implemented in the PPU of the TC49x microcontroller. The achieved execution time is below 4 µs, enabling the control of two 3-phase subsystems with a sampling period of 10 µs.
The autoencoder algorithm was implemented and tested on the PPU of the TC49x microcontroller as C code, utilizing vector functions from the Synopsys Vector DSP library. The implemented autoencoder fault detection algorithm was verified in both fault-free conditions and with inter-turn short circuit fault emulation (Figure 2). The computation time of the implemented autoencoder on the PPU for one three-phase subsystem is below 17 µs.
The operation of both proposed algorithms was verified on a laboratory testbench using a 2x 3-phase motor capable of emulating short circuit faults showing the benefit of PPU for parallel/AI computing.
|
|||
Pictures/visuals with titles
|
![]()
|