vikilee9/pcb-near-field-emi

PCB Near-Field EMI Classification Dataset

Printed Circuit Board near-field probe array measurements for electromagnetic interference (EMI) configuration classification.

Dataset Summary

This dataset contains CST full-wave electromagnetic simulation results of a 16-probe near-field array measuring the electromagnetic field above a PCB under 13 different design configurations. The task is to classify which PCB configuration produced a given near-field measurement—using only spatially-resolved field data, without access to the PCB schematic or S-parameters.

Key Features

• 16 probe positions in a 4×4 uniform grid (10mm spacing)
• 50 frequency points from 300–1000 MHz
• 6 complex field components per probe (Ex, Ey, Ez, real+imaginary)
• 13 PCB configurations (after S11-indistinguishable config merging)
• 10,400 total samples (6240 train, 4160 test)
• Frequency-blocked validation split built-in (train on low freqs, test on high freqs)

Why Frequency-Blocked?

Near-field EMI data exhibits strong frequency-domain autocorrelation (Pearson r=0.815). Random train/test splitting inflates accuracy by ~30 percentage points because the model can "cheat" by learning frequency-specific signatures rather than configuration-relevant features. This dataset enforces frequency-blocked splitting: training data comes from 300–714 MHz, test data from 729–1000 MHz. This is the only honest way to evaluate generalization.

Dataset Structure

pcb_nf_emi/
├── data/
│   ├── train.parquet    (6240 rows)
│   └── test.parquet     (4160 rows)
├── metadata.json
├── README.md
├── dataset_card.md
└── load_dataset.py

Features per row

Column	Type	Description
config_label	str	Merged configuration label (13 classes)
config_name	str	Original configuration name (16 classes)
probe	str	Probe identifier (P00-P33)
row	int	Grid row (0-3)
col	int	Grid column (0-3)
x_mm	float	Probe X position (mm)
y_mm	float	Probe Y position (mm)
z_mm	float	Probe Z position (mm, fixed 4.6)
frequency_mhz	float	Measurement frequency (MHz)
Ex_real/imag	float	X-component E-field (V/m)
Ey_real/imag	float	Y-component E-field (V/m)
Ez_real/imag	float	Z-component E-field (V/m)
E_abs	float	Total E-field magnitude

PCB Configurations

Label	Description	Count	Key Physical Signature
normal_power_equiv	Normal 50Ω termination	3 merged	Baseline, uniform field
gnd_slot	Ground plane slot	1	High center/edge E-field ratio
gnd_slot_narrow	Narrow ground slot	1	Similar to gnd_slot
impedance_stub_100ohm	100Ω stub	1	High impedance, strong reflection
impedance_stub_25ohm	25Ω stub	1	Low impedance
impedance_taper_25to50	25→50Ω taper	1	Impedance transition
mismatch_75ohm	75Ω mismatch	1	Moderate impedance mismatch
mismatch_100ohm_ref	100Ω mismatch	1	High impedance mismatch
near_trace_interfere	Probe near signal trace	2 merged	Enhanced trace-probe coupling
osc_xoffset_5mm	Oscillator X-offset 5mm	1	X-asymmetric field
osc_yoffset_5mm	Oscillator Y-offset 5mm	1	Y-asymmetric field, isolated cluster
port_100ohm	100Ω port	1	High-Z port
port_25ohm	25Ω port	1	Low-Z port

S11-Indistinguishable Merged Groups

The following configurations produce identical S11 (port reflection) signatures and are merged:

• normal_50ohm = power_variation = power_variation_half → normal_power_equiv
• near_trace_interfere ≈ near_trace_interfere_close → near_trace_interfere

These configurations require spatial near-field data for discrimination—S-parameter analysis alone cannot distinguish them.

Usage

import pandas as pd
import numpy as np

# Load data
train = pd.read_parquet("train.parquet")
test = pd.read_parquet("test.parquet")

# Build per-sample features: pivot (config, freq) → 16 probes × features
def build_samples(df):
    samples, labels, freqs = [], [], []
    for (config, freq), group in df.groupby(['config_label', 'frequency_mhz']):
        group = group.sort_values(['row', 'col'])
        features = group[['Ex_abs','Ey_abs','Ez_abs','E_abs',
                          'Ex_real','Ex_imag','Ey_real','Ey_imag',
                          'Ez_real','Ez_imag']].values  # (16, 10)
        samples.append(features)
        labels.append(config)
        freqs.append(freq)
    return np.stack(samples), np.array(labels), np.array(freqs)

X_train, y_train, f_train = build_samples(train)
X_test, y_test, f_test = build_samples(test)

print(f"Train: {X_train.shape}, Test: {X_test.shape}")

Benchmarks

Model	Features	Validation	Train Acc	Test Acc
RandomForest	16 (E_abs)	Random split	—	54.0% (inflated)
RandomForest	16 (E_abs)	Freq-blocked	—	~23%
CNN+Attention	96 (raw)	Freq-blocked	—	27.7%
LightGBM	70 (engineered)	Freq-blocked	—	50.0%
LightGBM+V3	100 (advanced)	4-fold freq-CV	—	53.9% ± 20.6%
Chance	—	—	—	7.7%

Key finding: Frequency-blocked validation reduces reported accuracy by ~30pp compared to random splitting. Always use the built-in train/test frequency split for honest evaluation.

DUT and Simulation Details

• Device Under Test: 369.5 MHz LC differential oscillator on 60×48mm, 4-layer FR4 PCB
• Solver: CST Studio Suite 2026.2, HF Time Domain (Hexahedral FIT)
• GPU: NVIDIA RTX 5080 (15.9 GB VRAM)
• Grid: 4×4 uniform, 10mm spacing, centered above oscillator
• Probe: Level 0 ideal E-field monitors (no physical probe perturbation)

Citation

@dataset{li2026pcbnfemi,
  title     = {PCB Near-Field EMI Classification Dataset},
  author    = {Li, Weiji},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/XXXX},
  note      = {Beihang University, School of Electronic and Information Engineering}
}

License

CC-BY-4.0 — Free to use with attribution.

Contact

Weiji Li (李炜基), Beihang University