Olfactory EEG Signal Classification Network

• The paper introduces OESCN, a deep learning framework that classifies olfactory EEG signals using adaptive frequency band extraction and subject-specific attention.
• It employs a four-stage pipeline—PSD estimation, multi-scale band generation, attention mechanism, and a spatio-spectral CNN—to process and classify EEG data.
• Benchmarking shows OESCN achieves 97.1% accuracy, outperforming EEGNet by 13.3% and reducing inter-subject variability significantly.

The Olfactory EEG Signal Classification Network (OESCN) is a deep learning framework specifically designed for classifying electroencephalogram (EEG) signals elicited by olfactory stimuli. The architecture emphasizes adaptive frequency band feature extraction and subject-specific spectral attention, coupled with a compact spatio-spectral convolutional neural network (CNN), achieving robust and high-accuracy classification across individuals for 13-class odor recognition tasks (Sun et al., 2022).

1. Architectural Overview

OESCN is constructed as a four-stage pipeline optimized for extracting discriminative representations from olfactory-induced EEG. The stages are:

1. Pre-processing & PSD Estimation: Raw EEG data $X \in \mathbb{R}^{C \times T}$ (C channels, T time samples) are processed via Welch’s periodogram per channel, producing a power spectral density (PSD) representation $F \in \mathbb{R}^{C \times P}$, where $P$ denotes frequency bins over 0.5–70 Hz.
2. Frequency Band Generator: A sliding-window mechanism extracts candidate frequency sub-bands over the PSD, aggregating multi-scale bandwise features into $S \in \mathbb{R}^{C \times K}$.
3. Frequency Band Attention Mechanism: Subject-specific attention is imposed on bandwise features: a multi-head self-attention incorporates a global and several local heads, followed by head fusion and a skip connection, yielding a re-weighted spatio-spectral map $M' \in \mathbb{R}^{C \times K}$.
4. Spatio-Spectral CNN Classifier: $M'$ serves as a single-channel 2D “image,” processed via parallel convolutions of varying kernel sizes, pooled, and passed through fully-connected (FC) layers to output a 13-way softmax for odor identification.

The data-flow is: $$\text{EEG input}~ X \rightarrow \text{Welch PSD} \rightarrow \text{Band Generator} \rightarrow \text{Attention} \rightarrow \text{CNN} \rightarrow \text{softmax}$$.

2. Frequency Band Generation

The frequency band generator performs an exhaustive, multi-scale sweep over the PSD for each channel:

• For window lengths $L_i \in \{1, 5, 10, 15, 20\}$ Hz, with $G = 1$ Hz step, the number of bands per $L_i$:

$F \in \mathbb{R}^{C \times P}$0

• For each slice $F \in \mathbb{R}^{C \times P}$1, average spectral power over $F \in \mathbb{R}^{C \times P}$2:

$F \in \mathbb{R}^{C \times P}$3

• Concatenate all $F \in \mathbb{R}^{C \times P}$4 to form $F \in \mathbb{R}^{C \times P}$5, $F \in \mathbb{R}^{C \times P}$6.
• Band-combination tensor $F \in \mathbb{R}^{C \times P}$7 is obtained by stacking across channels.

This process enables high-resolution, adaptive capturing of both narrow and broad frequency features, essential for encoding the olfactory event-related EEG.

3. Frequency Band Attention Mechanism

This module adaptively emphasizes subject-relevant bands through multi-head self-attention:

• Global Head:
  • Linear transformations:

  $F \in \mathbb{R}^{C \times P}$8

  where $F \in \mathbb{R}^{C \times P}$9. - Scaled dot-product attention:

  $P$0

• Local Heads (for each $P$1): Apply attention as above to each $P$2.

• Head Fusion:

  • Concatenate outputs: $P$3.
  • Max-pooling and average-pooling:

  $P$4 - Fuse via $P$5 convolution:

  $P$6 - Add skip connection:

  $P$7

The attention mechanism is explicitly trained for each subject’s dataset split, ensuring adaptation to inter-subject spectral variability.

4. Spatio-Spectral CNN Classifier

The CNN receives $P$8 interpreted as a single-channel image of size $P$9:

• Layer 0: Reshape to $S \in \mathbb{R}^{C \times K}$0.

• Layer 1: Parallel 2D convolutions:

  • $S \in \mathbb{R}^{C \times K}$1, $S \in \mathbb{R}^{C \times K}$2, and $S \in \mathbb{R}^{C \times K}$3 kernels ($S \in \mathbb{R}^{C \times K}$4 filters each), ELU activations, concatenated → $S \in \mathbb{R}^{C \times K}$5.
• Layer 2: Average-pooling across spatial dimensions.
• Layer 3: $S \in \mathbb{R}^{C \times K}$6 conv with $S \in \mathbb{R}^{C \times K}$7 filters (ELU), average-pooling.
• Layer 4: Flatten, FC (128 units, ELU, BN, Dropout 0.25), FC (64 units, ELU, BN, Dropout 0.25), FC (13 units, softmax).

Typical hyperparameters: $S \in \mathbb{R}^{C \times K}$8, $S \in \mathbb{R}^{C \times K}$9, stride=1, padding="same". This structure exploits electrode (spatial) and frequency-band (spectral) topology.

5. Training and Evaluation Protocol

• Dataset: 11 healthy subjects, 13 odors, 35 trials each, yielding 5,005 total trials. EEGs are sampled at 1 kHz over 32 channels (30 analyzed), 10 seconds per trial.
• Pre-processing: PSD extracted via Welch’s method (Hamming window, 200 samples, overlap 8).
• Cross-validation: 10-fold, per subject.
• Optimization: Cross-entropy loss, Adam ($M' \in \mathbb{R}^{C \times K}$0), 500 epochs, batch size 39.

Performance Benchmarking:

OESCN was benchmarked against EEGNet and AFBD-SVM; results are summarized below:

Method	Average Accuracy (%)	Inter-subject Std
EEGNet	83.8	12.3
AFBD-SVM	87.3	10.2
OESCN	97.1	3.7

OESCN delivers a 13.3% gain over EEGNet and substantially reduces inter-subject variability.

6. Ablation Analysis

Two OESCN variants were constructed to assess contribution of each module:

• OESCN_a1: removes the attention mechanism, passing $M' \in \mathbb{R}^{C \times K}$1 directly to the CNN.
• OESCN_a2: removes both band generator and attention, using fixed uniform sub-bands plus CNN.

Variant	Avg Acc (%)	Inter-subj Std
OESCN_a2	94.3	5.3
OESCN_a1	95.9	4.5
OESCN	97.1	3.7

Removing attention reduces accuracy by ≈1.2%. Eliminating both generator and attention further reduces accuracy by ≈1.6%. This demonstrates that both exhaustive band extraction and subject-specific attention are critical for top performance and robustness.

7. Algorithmic Workflow

A single training epoch for one subject proceeds as:

$M' \in \mathbb{R}^{C \times K}$2

The design leverages a comprehensive multiscale frequency extraction strategy and lightweight subject-specific attention. This hybrid architecture yields state-of-the-art olfactory EEG classification accuracy and significantly enhanced inter-subject robustness (Sun et al., 2022).