On-Target Embedded Benchmarking of Mamdani and Neural-Network-Tuned Fuzzy Controllers for Automotive Deceleration Command Computation

Intelligent braking controllers are usually evaluated in simulation or by vehicle-level performance, but few studies isolate the embedded computational cost of deploying them on the same microcontroller. This paper presents a same-platform benchmarking framework running a baseline Mamdani fuzzy logic controller and a neural-network-tuned variant on one STM32H723ZG target, sharing a common CAN input, a 27-rule Mamdani inference path, and an on-target Cortex-M7 cycle-counter (DWT) timing method that needs no external hardware. The controllers differ only in a compact feedforward network that adjusts three speed and three distance membership-function breakpoints, isolating the tuning layer’s cost. A 450-case zero-tuning test reproduces the baseline exactly, and a 443-vector Colab-to-STM32 campaign confirms the deployed network matches its offline reference to within 2.38 × 10−7. Over nine scenarios and 9000 samples, in an optimization-enabled (–O2) build, the tuning layer added 10.5 µs (6.29%) and under 1% of the 20 ms control period, with no observed controller-execution deadline miss. A Float32 sweep of 886 operating points shows the layer moves the deceleration command by up to 1.332 m/s2, concentrated where the membership functions are unsaturated. The contribution is a reproducible methodology quantifying the on-target execution cost, deployment fidelity and functional authority of a neural-network tuning layer on a high-performance Cortex-M7 embedded microcontroller.