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 $L_i$:

$$B_{ci} = \left\lfloor \frac{P - L_i}{G} \right\rfloor$$

• For each slice $$A^{c, i} \in \mathbb{R}^{1 \times L_i \times B_{ci}}$$, average spectral power over $L_i$:

$$D^{c, i}_j = \frac{1}{L_i} \sum_{k=1}^{L_i} A^{c, i}_{k, j} \quad \text{for}~ j = 1, ..., B_{ci}$$

• Concatenate all $D^{c, i}$ to form $S_c \in \mathbb{R}^{1 \times K}$, $K = \sum_{i=1}^5 B_{ci}$.
• Band-combination tensor $S \in \mathbb{R}^{C \times K}$ 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:

  $$Q^{\mathrm{glo}} = S W^q ,~~ K^{\mathrm{glo}} = S W^k ,~~ V^{\mathrm{glo}} = S W^v$$

  where $W^{(\cdot)} \in \mathbb{R}^{K \times K}$. - Scaled dot-product attention:

  $$H^{\mathrm{glo}} = \mathrm{Softmax}\left(\frac{Q^{\mathrm{glo}} (K^{\mathrm{glo}})^\top}{\sqrt{C}}\right) V^{\mathrm{glo}}~\in~\mathbb{R}^{C \times K}$$

• Local Heads (for each $L_i$): Apply attention as above to each $S^i \in \mathbb{R}^{C \times B_{ci}}$.

• Head Fusion:

  • Concatenate outputs: $$X = \mathrm{Concat}(H^{\mathrm{glo}}, H^{\mathrm{loc}})$$.
  • Max-pooling and average-pooling:

  $$X_{\max} = \mathrm{MaxPool}(X), \qquad X_{\mathrm{avg}} = \mathrm{AvgPool}(X)$$ - Fuse via $1\times1$ convolution:

  $$M = \mathrm{Conv}_{1\times1}\left( \mathrm{Concat}(X_{\max}, X_{\mathrm{avg}}) \right)$$ - Add skip connection:

  $M' = M + S$

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 $M' \in \mathbb{R}^{C \times K}$ interpreted as a single-channel image of size $(C, K)$:

• Layer 0: Reshape to $(1, C, K)$.

• Layer 1: Parallel 2D convolutions:

  • $3\times3$, $8\times8$, and $15\times15$ kernels ($F_1$ filters each), ELU activations, concatenated → $(3F_1, C, K)$.
• Layer 2: Average-pooling across spatial dimensions.
• Layer 3: $3\times3$ conv with $F_2$ 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: $F_1=32$, $F_2=64$, 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 ($1\times10^{-4}$), 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 $S$ 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:

for each minibatch {X, y}: # 1. Compute PSD F = Welch_PSD(X) # shape (batch, C, P) # 2. Band generator S = [] for i in 1..5: Li = {1,5,10,15,20}[i] slide_i = sliding_windows(F, length=Li, step=1) D_i = mean_over_window(slide_i) # shape (batch, C, B_i) S.append(D_i) S = concat_along_band_dim(S) # shape (batch, C, K) # 3. Attention H_glo = SelfAttention(S) # global head H_loc = concat(SelfAttention(splits of S)) # local heads M = Conv1x1(concat(maxpool,[H_glo,H_loc], avgpool,[H_glo,H_loc])) M_prime = M + S # 4. CNN Classifier logits = CNN(M_prime) # shape (batch, 13) loss = CrossEntropy(logits, y) backpropagate & update params

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).