# Machine Learning-Driven Sensitivity Analysis for a 2-Layer Printed Circuit Board Inductive Motor Position Sensor

## TL;DR

This paper introduces a machine learning approach to analyze and optimize a compact inductive motor position sensor design for automotive applications.

## Contribution

A novel ML-driven sensitivity analysis workflow for optimizing inductive sensor layouts and determining installation tolerances.

## Key findings

- The sensor accuracy remained within ±1 electrical degree across all test conditions.
- Tilt and Y-offset were identified as the main error contributors with strong interaction.
- XGBoost outperformed linear regression, showing non-linear accuracy degradation.

## Abstract

Motor position sensors are critical parts for traction motors control in electrified automotive powertrains. As motors are becoming more compact due to the advance of technology the packaging space for motor position sensors is becoming increasingly restricted. This study presents a two-layer (2L) printed circuit board (PCB) routing strategy for inductive motor position sensors with limited area. A prototype was fabricated and tested on a test bench using a comprehensive design of experiments that contains 625 combinations of X- and Y-offsets, tilt angle, and airgap at various levels (±0.5 mm in X/Y, ±0.5° tilt, 1.9–3.1 mm airgap). Across the tolerance box, the accuracy under all test cases remained within ±1 electrical degree. The accuracy analysis through Fourier series on a circle shows that the DC offset and magnitude mismatches of the 3 Rx signals are the dominant error contributors due to the routing modification. An Extreme Gradient Boosting (XGBoost) model was trained and validated with R2 = 0.9951. A comparison with a Multiple Linear Regression baseline (R2 = 0.0565) demonstrates that installation-induced accuracy degradation is inherently non-linear. The SHapley Additive exPlanations (SHAP) and interaction intensity analysis identified tilt and Y-offset as dominant error drivers, revealing a strong coupled influence (interaction intensity = 0.9581). The model revealed a mild Y-axis asymmetry introduced by routing modifications. This integrated workflow provides a general, quantitative framework for optimizing and analyzing inductive sensor layouts and establishing installation tolerances.