ORAL SESSION

Paper ID: IECON25-001668

Paper Title: GAN-Driven Signal Denoising and Enhancement for Robust Drone Motor Detection

Author’s Name: Dilshara Herath, Chinthaka Abeyrathne, Supun Ganegoda, Chatura Seneviratne, Harindra S. Mavikumbure

Department, Affiliation, Country: Dept. of Electrical and Information Engineering, University of Ruhuna, Dept. of Computer Sciences, Virginia Commonwealth University

Introduction

• Detecting Micro-Unmanned Aerial Systems (Micro-UAS) has become vital with its widespread use.
• Drone detection and motor fault diagnosis heavily rely on the analysis of electromagnetic signals emitted by BLDC motors. However, these signals are often obscured by low signal-to-noise ratios (SNR) due to environmental interference.

Generative Adversarial Networks for High-Fidelity Signal Denoising in Drone System Motors

System Architecture

Dataset

GAN-based Architecture for Signal Denoising

​

Generator

CNN consisting 3 key stages,

• Downsampling Stage: An initial 1D convolutional layer with 64 filters, a kernel size of 25, and a stride of 4 reduces the temporal dimension of the input while extracting salient features. This is followed by a Leaky ReLU activation with a slope coefficient of 0.2.
• Residual Processing: A sequence of five residual blocks processes the downsampled features. Each block includes two 1D convolutional layers with 64 filters and a kernel size of 9, augmented by batch normalization and Leaky ReLU activations.
• Upsampling Stage: A 1D transposed convolutional layer with a single filter, a kernel size of 25, and a stride of 4 reconstructs the signal to its original length.

Discriminator

• Architecture consists of a series of 1D convolutional layers with progressively increasing filter counts (64, 128, 256, 512), each employing a kernel size of 15 and a stride of 2 for hierarchical feature extraction.
• Batch normalization and Leaky ReLU activations (slope 0.2) follow all but the first convolutional layer to stabilize training and enhance gradient flow. The resulting feature maps are
• Flattened and fed into a dense layer with a sigmoid activation.

Why SCF ?

• Cyclostationarity occurs when statistical properties of a signal exhibit periodicity

Cyclostationary Property Analysis

CNN Model for Evaluation

CNN Model for Evaluation

GAN Model Performance

Boxplots for performance parameters

Evaluation by CNN Model

Conclusion

Team

Thank You!

Q/A