STFT-AECNN: Attn-Enhanced CNN for DAS

• The paper presents STFT-AECNN, which transforms multi-channel raw signals into stacked spectrogram tensors and classifies events using a lightweight, attention-augmented hierarchical CNN.
• It integrates a Spatial Efficient Attention Module (SEAM) to recalibrate channel features, yielding superior accuracy (99.94%) and enhanced efficiency with only 0.43M parameters.
• The model’s joint Cross-Entropy and Triplet loss optimizes both class boundaries and intra-class compactness, demonstrating robust and fine-grained event recognition in DAS.

Searching arXiv for the exact STFT-AECNN paper and closely related STFT+attention/CNN work to ground the article in current literature. STFT-based Attention-Enhanced Convolutional Neural Network (STFT-AECNN) denotes a neural architecture in which short-time Fourier transform (STFT) converts multi-channel time-series signals into spectrogram tensors and a convolutional backbone, augmented by lightweight attention, performs classification on the resulting time-frequency representation. In the literature, the exact designation STFT-AECNN refers to a model proposed for event recognition in phase-sensitive optical time-domain reflectometry ($\Phi$-OTDR) within distributed acoustic sensing (DAS), where per-channel spectrograms are stacked and processed by a hierarchical 2D CNN with a Spatial Efficient Attention Module (SEAM) and a joint Cross-Entropy plus Triplet objective (Lan et al., 23 Sep 2025). Related work in other domains has used closely allied design patterns—STFT front ends, convolutional backbones, and explicit or implicit time-frequency selection mechanisms—for speech enhancement, EEG decoding, image recognition, and biomedical audio classification (Wan et al., 2023, Ali et al., 2020, Kumawat et al., 2020, Chen et al., 2022).

1. Definition and conceptual basis

The STFT-AECNN formulation in the strict sense was introduced for event recognition in phase-sensitive optical time-domain reflectometry ($\Phi$-OTDR), a sensing modality used in distributed acoustic sensing (DAS) for applications including smart-city surveillance, industrial pipeline monitoring, and critical infrastructure protection (Lan et al., 23 Sep 2025). Its central premise is that $\Phi$-OTDR data are inherently multi-channel spatiotemporal signals whose discriminative information is distributed across time, frequency, and space. The model therefore avoids both raw-sequence flattening and crude image conversion, instead transforming each spatial channel into a spectrogram and stacking these spectrograms into a tensor suitable for efficient 2D CNN processing (Lan et al., 23 Sep 2025).

This design addresses three limitations identified in earlier approaches: loss of spatial structure in 1D sequential models, suboptimal image representations such as grayscale mappings or Mel spectrograms for physical vibration signals, and the accuracy–efficiency trade-off of heavy LSTM or Transformer architectures in resource-constrained IoT edge deployment (Lan et al., 23 Sep 2025). The resulting method is explicitly intended to preserve the multi-channel spatiotemporal structure, exploit a high-fidelity time-frequency representation, retain the efficiency of 2D CNNs, and use lightweight attention to emphasize informative channels (Lan et al., 23 Sep 2025).

A broader reading of the literature shows that this architectural pattern is not unique to DAS. In speech enhancement, the model termed MNTFA operates on STFT-domain inputs, uses a convolutional encoder-decoder, inserts axial self-attention over time and frequency, predicts a complex ratio mask, and reconstructs via iSTFT (Wan et al., 2023). In EEG decoding, anchored-STFT produces multiple spectrogram resolutions and a CNN classifier performs recognition, although no explicit attention module is used (Ali et al., 2020). These parallel developments suggest that STFT-AECNN is both a named model class in DAS and a more general architectural idiom in time-frequency deep learning.

2. STFT representation and stacked spectrogram input

In the canonical STFT-AECNN design, the raw event sample is a matrix of dimension

$10000 \times 12,$

corresponding to $10000$ temporal samples and $12$ adjacent spatial channels (Lan et al., 23 Sep 2025). The paper defines raw data as

$\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$

where $N_s$ is the number of spatial channels and $N_t$ is the number of time samples (Lan et al., 23 Sep 2025). For each channel $i$, the raw 1D signal is transformed by STFT into a time-frequency matrix. The continuous and discrete formulations are given as

$\Phi$0

and

$\Phi$1

where $\Phi$2 is the window length, $\Phi$3 the hop length, $\Phi$4 the FFT size, $\Phi$5 the frame index, and $\Phi$6 the frequency-bin index (Lan et al., 23 Sep 2025).

The magnitude spectrogram is defined as

$\Phi$7

After magnitude extraction, the pipeline applies logarithmic compression $\Phi$8 and min-max normalization to $\Phi$9 (Lan et al., 23 Sep 2025). Each processed channel becomes a spectrogram $\Phi$0, and the final stacked input tensor is

$\Phi$1

In the reported experiments, this becomes

$\Phi$2

so each event is represented as a 12-channel spectrogram tensor, with each channel a $\Phi$3 spectrogram (Lan et al., 23 Sep 2025).

The STFT preprocessing parameters are reported explicitly. For each of the 12 channels, the input length per channel is $\Phi$4; the window type is Hann; the window length is $\Phi$5; the FFT size is $\Phi$6; the hop length is $\Phi$7; the overlap length is $\Phi$8; and the target spectrogram size is $\Phi$9 (Lan et al., 23 Sep 2025). This representation preserves time-frequency structure inside the spectrogram plane and spatial identity across channels, allowing the network to learn local patterns such as impulsive bursts, sustained oscillatory bands, broadband textures, and smooth low-frequency trajectories while still reasoning over cross-channel spatial correlations (Lan et al., 23 Sep 2025).

The paper motivates this representation with a concrete example: in a watering sample, discriminative patterns are concentrated in channels 4–8, while other channels mostly contain noise (Lan et al., 23 Sep 2025). This is a central empirical rationale for channel-selective attention rather than uniform treatment of all stacked spectrogram channels.

3. Architecture: hierarchical 2D CNN with SEAM

The STFT-AECNN pipeline consists of raw multi-channel $10000 \times 12,$0-OTDR time-series input, per-channel STFT preprocessing, stacking of spectrograms into a multi-channel tensor, a hierarchical 2D CNN backbone, SEAM attention in each stage, global pooling and a classifier head, and joint optimization with Cross-Entropy plus Triplet loss (Lan et al., 23 Sep 2025). The architecture is described as a hierarchical four-stage architecture, where each stage contains a CBRP block, a split into a SEAM attention path and a residual identity path, and a fusion step with learnable scalar $10000 \times 12,$1 (Lan et al., 23 Sep 2025).

CBRP denotes Convolution, Batch Normalization, ReLU, and Max Pooling (Lan et al., 23 Sep 2025). The $10000 \times 12,$2-th output channel of the convolution in the CBRP block is written as

$10000 \times 12,$3

BatchNorm and ReLU follow: $10000 \times 12,$4

$10000 \times 12,$5

and after max pooling,

$10000 \times 12,$6

The paper does not provide exact numerical values for convolution kernel sizes, output channel counts per stage, pooling kernel sizes or strides, dropout rate, or hidden dimensions in the classifier head (Lan et al., 23 Sep 2025).

The Spatial Efficient Attention Module (SEAM) is the model’s defining attention mechanism. Although named “spatial,” operationally it is a lightweight channel attention module driven by global average pooling and global max pooling of the CBRP output (Lan et al., 23 Sep 2025). Its adaptive 1D convolution kernel size is

$10000 \times 12,$7

with typical defaults

$10000 \times 12,$8

The notation $10000 \times 12,$9 denotes the nearest odd integer (Lan et al., 23 Sep 2025). The attention weights are then

$10000$0

and the recalibrated feature map is

$10000$1

Stage-level fusion combines the attended and original features as

$10000$2

or equivalently,

$10000$3

This residual-style fusion ensures that attended features do not wholly replace the baseline signal path (Lan et al., 23 Sep 2025).

After the four stages, the network applies Global Average Pooling, flattening, a first fully connected layer, ReLU, Dropout, and a second fully connected layer producing logits for $10000$4 classes (Lan et al., 23 Sep 2025). The embedding used for metric learning is extracted from the backbone before the final classification layers, but the exact embedding dimensionality is not provided (Lan et al., 23 Sep 2025). The model is explicitly designed to remain lightweight: it uses a 2D CNN rather than LSTM or Transformer, a single-stream backbone, lightweight attention via SEAM, and has 0.43M parameters in the reported results (Lan et al., 23 Sep 2025).

4. Objective functions and discriminative learning

STFT-AECNN is trained with a joint Cross-Entropy and Triplet loss. The Cross-Entropy term is

$10000$5

where $10000$6 is the logit vector and $10000$7 is the ground-truth class index (Lan et al., 23 Sep 2025). The Triplet loss acts on embeddings $10000$8, $10000$9, and $12$0 for anchor, positive, and negative samples: $12$1 where $12$2 is Euclidean distance and $12$3 is the margin (Lan et al., 23 Sep 2025). The total objective is an unweighted sum,

$12$4

The paper’s motivation for this dual objective is that Cross-Entropy alone optimizes class boundaries but does not explicitly impose intra-class compactness and inter-class separation in the embedding space, which is particularly important for subtle or fine-grained $12$5-OTDR events such as watering versus walking (Lan et al., 23 Sep 2025). The reported training protocol sets the triplet margin to $12$6 and uses Adam, initial learning rate $12$7, weight decay $12$8, 50 epochs, batch size 50, and a fixed random seed (Lan et al., 23 Sep 2025). The implementation environment is also reported in detail: Python 3.8 on Ubuntu 20.04, PyTorch 1.10.0, CUDA 11.3, NVIDIA GeForce RTX 4090 with 24 GB, Intel Xeon Gold 6430, and 120 GB RAM (Lan et al., 23 Sep 2025).

This coupling of a lightweight channel-attentive CNN with metric learning distinguishes STFT-AECNN from several earlier STFT-based CNNs that rely solely on standard classification loss. In lung-sound classification, for example, a fine-tuned ResNet18 with STFT features uses a weighted loss but no metric-learning term (Chen et al., 2022). In speech enhancement, MNTFA couples STFT-domain convolution and time-frequency attention with a multi-loss objective, but the losses are spectrogram MSE, multi-resolution STFT, and WavLM-based ASR loss rather than Cross-Entropy plus Triplet (Wan et al., 2023). This suggests that the exact loss design in STFT-AECNN is closely tied to discriminative event recognition rather than signal reconstruction.

5. Experimental protocol and empirical performance

The experiments use the public BJTU $12$9-OTDR dataset from Cao et al. (Lan et al., 23 Sep 2025). The acquisition setup uses an intensity-based $\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$0-OTDR system with sensing fiber lengths of 5 km and 10 km, both terminated by a 100 m armored section, with events induced within the first 50 m of the armored fiber (Lan et al., 23 Sep 2025). Pulse repetition frequency is 12.5 kHz for 5 km and 8 kHz for 10 km, while DAQ sampling rate is 10 MSa/s in both settings (Lan et al., 23 Sep 2025).

The dataset contains 15,419 samples across six classes: background noise (2946), digging (2512), knocking (2530), watering (2253), shaking (2728), and walking (2450) (Lan et al., 23 Sep 2025). Each sample has shape

$\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$1

and the train/test protocol follows an 8:2 random split with no overlap between splits (Lan et al., 23 Sep 2025). The evaluation metrics are overall Accuracy and per-class Precision, Recall, and F1-score, defined by the standard formulas

$\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$2

$\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$3

$\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$4

$\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$5

The headline result is a peak accuracy of $\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$6 (Lan et al., 23 Sep 2025). Per-class results are highly balanced: background and digging both reach Precision 100.0, Recall 100.0, F1 100.0; knocking reaches Precision 100.0, Recall 99.8, F1 99.9; watering reaches Precision 99.8, Recall 99.8, F1 99.8; shaking reaches Precision 100.0, Recall 100.0, F1 100.0; walking reaches Precision 99.8, Recall 100.0, F1 99.9; and the averages for Precision, Recall, and F1 are all 99.9 (Lan et al., 23 Sep 2025).

Comparison against baselines shows STFT-AECNN: 99.94%, ViT-VSEC: 98.99%, ST-T: 97.99%, 2D-CNN: 93.74%, and SVM: 89.88% (Lan et al., 23 Sep 2025). The paper further reports that the full model outperforms two ablated variants: w/o SEAM Attention: 99.77% and w/o Triplet Loss: 99.81%, corresponding to drops of 0.17% and 0.13% respectively (Lan et al., 23 Sep 2025). The confusion matrix is described as almost perfectly diagonal, with only two misclassifications: one knocking sample misclassified as watering, and one watering sample misclassified as walking (Lan et al., 23 Sep 2025). The t-SNE analysis also indicates more compact and better-separated class clusters than SVM, 2D-CNN, ST-T, and ViT-VSEC (Lan et al., 23 Sep 2025).

Efficiency claims are central to the STFT-AECNN formulation. Reported inference speed is 0.35 ms/sample for STFT-AECNN, compared with 1.86 ms/sample for ST-T and 2.82 ms/sample for ViT-VSEC (Lan et al., 23 Sep 2025). Model size is 0.43M parameters, approximately 10% of the Transformer models, and FLOPs are reported as less than 20% of ST-T and less than 10% of ViT-VSEC (Lan et al., 23 Sep 2025). Training time per epoch is said to be only 10% of ViT-VSEC and 27% of ST-T, while even surpassing the baseline 2D-CNN in training speed (Lan et al., 23 Sep 2025). These empirical results underpin the paper’s claim that a carefully designed STFT-based CNN with lightweight attention can deliver both near-perfect recognition and edge-oriented efficiency.

6. Broader research context, interpretations, and limitations

Within the broader STFT-CNN literature, STFT-AECNN sits at the intersection of three lines of work: STFT as a front-end representation, CNNs as efficient local pattern extractors, and attention or multi-resolution mechanisms as a means of selecting informative time-frequency structure.

One relevant precursor is the use of anchored-STFT in EEG motor-imagery decoding, where multiple anchor lengths $\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$7 generate multiple spectrogram views and a shallow CNN named Skip-Net classifies each view, followed by majority voting (Ali et al., 2020). That system achieves average classification accuracy 89.5% on BCI Competition IV dataset 2b and 90.7% on BCI Competition II dataset III, and its ablation shows that Anchored-STFT + Skip-Net outperformed STFT by 3.7% and CWT by 3.6% on one evaluation setting (Ali et al., 2020). The model, however, uses no explicit attention mechanism, no self-attention, no channel attention, and no temporal attention (Ali et al., 2020). A plausible implication is that STFT-AECNN can be understood as a more integrated successor to such pipelines: rather than performing resolution-level voting or manual fusion, it learns channel weighting directly within the convolutional backbone.

In speech enhancement, MNTFA provides a second instructive comparison. It operates explicitly in the STFT domain, uses a convolutional encoder-decoder, inserts axial self-attention over time and frequency, predicts a complex ratio mask, and reconstructs through iSTFT (Wan et al., 2023). The model has 0.23M parameters, 1.89 GFLOPS, and outperforms DPCRN in PESQ, STOI, and WER on the reported DNS setup (Wan et al., 2023). This indicates that “STFT-based attention-enhanced CNN” is not confined to DAS; it also describes a broader family of time-frequency architectures where attention is used to model long-range temporal or spectral dependencies.

Further context comes from STFT-based convolutional operators that omit explicit attention. Depthwise-STFT based separable Convolutional Neural Networks replace trainable depthwise spatial convolution with local Fourier responses at fixed low-frequency points and linear pointwise mixing, outperforming standard depthwise separable baselines on CIFAR-10 and CIFAR-100 with reduced space-time complexity (Kumawat et al., 2020). Depthwise Spatio-Temporal STFT Convolutional Neural Networks for Human Action Recognition replace 3D convolutional layers with fixed STFT filtering plus trainable channel mixing, using roughly 3.5 to 4.5 times fewer parameters and 1.5 to 1.8 times fewer computations than the compared state-of-the-art methods (Kumawat et al., 2020). These results suggest that STFT-based convolution can itself act as a strong inductive bias even before attention is added.

STFT-AECNN’s reported limitations are mostly implicit rather than extensively analyzed. The experiments are conducted on the BJTU dataset only, without cross-dataset validation or domain-shift analysis, and the accessible text omits exact kernel sizes per stage, channel counts per stage, exact hidden dimensions, dropout rate, and triplet mining strategy (Lan et al., 23 Sep 2025). The runtime measurements are on an RTX 4090 server rather than embedded hardware, even though the model is positioned as edge-friendly (Lan et al., 23 Sep 2025). This suggests that the model’s deployment claims are strongest at the level of comparative efficiency rather than demonstrated embedded implementation.

A final interpretive point concerns STFT itself. In monaural speech separation, a direct front-end comparison found that CQT-based systems consistently outperform STFT-based systems, with an average gain of about 0.4 dB SDRi across tested back ends on WSJ0-2mix (Shi et al., 2019). That study does not imply that STFT is unusable; rather, it positions STFT as strong and practical but slightly suboptimal because of its linear frequency spacing (Shi et al., 2019). This suggests that STFT-AECNN’s success should not be read as a general proof of STFT optimality across all domains, but as evidence that in $\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$8-OTDR event recognition, a well-chosen STFT representation aligns effectively with both the signal structure and the computational constraints of the task.

In summary, STFT-AECNN is best understood as a compact, task-specific synthesis of STFT preprocessing, stacked multi-channel spectrogram representation, hierarchical 2D convolution, lightweight channel-selective attention, and metric-aware supervision. Its defining contribution is not merely the use of STFT or CNNs in isolation, but the demonstration that this combination can preserve spatiotemporal structure, adaptively emphasize informative channels, and achieve 99.94% peak accuracy with 0.43M parameters and 0.35 ms/sample inference on a public $\mathbf{X}_{\text{raw} \in \mathbb{R}^{N_s \times N_t},$9-OTDR benchmark (Lan et al., 23 Sep 2025).