Dilated CNNs for Periodic Signal Processing: A Low-Complexity Approach

1. Problem Statement

The paper addresses the challenge of denoising periodic signals (e.g., speech, ECG/EEG, radar, sonar) in environments with strict power and computational constraints, such as IoT and edge devices.

• Current Limitations:
  • Classical Methods: Techniques like Autoregressive (AR) modeling often require long, stationary recordings to achieve high frequency resolution and may struggle with varying fundamental frequencies.
  • Standard Deep Learning: While Deep Convolutional Neural Networks (DCNNs) offer superior performance, they typically require re-training for every new signal observation. This per-observation training incurs massive computational overhead, making them unsuitable for resource-constrained edge devices.
  • Fixed Dilation Constraints: Standard 1D Dilated CNNs use fixed integer dilation factors (e.g., $2^{l-1}$). This works well for signals with constant periods but fails when the signal's fundamental frequency ($F_0$) varies, as the temporal alignment of the dilation kernels becomes misaligned with the signal's periodicity.

Goal: Develop a method that achieves denoising accuracy comparable to state-of-the-art deep learning and classical methods but with minimal computational complexity and no re-training for new observations, provided the fundamental frequency is known.

2. Methodology: R-DCNN

The authors propose R-DCNN (Resampling-based Dilated CNN), a two-stage architecture that decouples training from inference via time-scale alignment.

Core Concept

The method relies on the mathematical insight that if a periodic signal is resampled (time-scaled) to match a reference frequency, the same set of learned weights can be applied to predict the signal, regardless of the original frequency.

The Two Stages

1. Training Stage (Single Observation):

   • A 1D linear Dilated CNN is trained on one single reference observation with a known fundamental frequency ($F_{ref}$).
   • The network learns optimal weights ($W$) to predict the deterministic periodic component from the noisy input.
   • The architecture uses dilated convolutions to expand the receptive field without increasing parameters or losing resolution.

2. Inference Stage (New Observations):

   • For a new observation with a different fundamental frequency ($F$), the system calculates a scaling factor $s = F / F_{ref}$.
   • Resampling: The new noisy signal is resampled (interpolated) to align its time axis with the reference signal. This involves upsampling and downsampling using rational factors ($U/L$) to approximate $s$ efficiently (using polyphase filtering).
   • Prediction: The fixed, pre-trained weights are applied to the resampled signal to estimate the clean waveform. No re-training occurs.

Mathematical Formulation

If the reference signal predictor is $\hat{x}_{ref}[n] = \sum w_m \cdot y_{ref}[n-m]$, then for a signal with frequency scaled by $s$:
$$\hat{x}_i[n/s] = \sum w_m \cdot \tilde{y}_i[n/s - m/s]$$
Where $\tilde{y}$ is the interpolated observation. The weights $w_m$ remain invariant; only the time grid changes.

3. Key Contributions

• Single-Training, Multi-Frequency Inference: The primary contribution is a framework where a DCNN is trained once on a single observation and generalized to signals with varying fundamental frequencies via a lightweight resampling step.
• Low-Complexity Edge Deployment: By eliminating the need for per-observation retraining, the method drastically reduces computational load and energy consumption, making it viable for IoT devices.
• Theoretical Justification: The paper proves (via Wold's theorem and linear prediction theory) that for deterministic periodic processes, the optimal predictor coefficients are unique and invariant to time scaling, provided the scaling factor is known.
• Efficient Resampling: The authors propose using rational approximations and polyphase filtering to implement the time-scaling step with minimal computational cost.

4. Experimental Results

The authors evaluated R-DCNN against two baselines:

1. Conventional DCNN: Retrained for every single observation.
2. Autoregressive (AR) Modeling: Optimized per observation.

Experimental Setup:

• Signals: Sum of sinusoids with fundamental frequencies ranging from 20–50 Hz and varying numbers of harmonics (5, 10, 20).
• Noise: Additive white noise with SNR ranging from -10 dB to 15 dB.
• Metric: Output Signal-to-Noise Ratio (SNR) after denoising.

Findings:

• Performance Parity: R-DCNN achieved denoising performance comparable to the re-trained DCNN and significantly better than AR methods, particularly at lower SNR levels.
• Efficiency: While the re-trained DCNN offered slightly better peak performance, it required training for every signal. R-DCNN achieved similar average performance with zero training cost during inference.
• Robustness: The method remained stable across different network depths (number of layers). However, performance degraded if the receptive field (determined by layer count and kernel size) was too small to capture the necessary frequency resolution.
• Waveform Estimation: Visual inspection of waveforms at various SNRs (-5 dB to 5 dB) confirmed that R-DCNN effectively recovered the underlying periodic structure even in high-noise environments.

5. Significance and Future Work

Significance:
This work bridges the gap between high-performance deep learning and the strict resource constraints of edge computing. It demonstrates that frequency adaptability in deep learning does not necessarily require retraining or complex dynamic architectures; a simple time-scaling (resampling) mechanism is sufficient to generalize a fixed-weight model across varying periodic signals.

Limitations & Future Directions:

• Frequency Estimation Dependency: The current method assumes the fundamental frequency ($F_0$) of the new signal is known. In practice, estimating $F_0$ accurately in real-time adds computational overhead.
• Future Work: The authors plan to investigate the robustness of R-DCNN against errors in fundamental frequency estimation and to integrate low-complexity frequency estimation techniques directly into the pipeline.

In conclusion, R-DCNN offers a practical, efficient solution for periodic signal denoising in IoT and medical devices, maintaining high accuracy while adhering to strict power and memory budgets.