Time–Frequency-Based CNN Waveform Classification: Comparison of Time-Domain, FFT, STFT, Wavelet and Hybrid Deep Learning

Reliable classification of noisy waveforms is an important problem in modern signal processing, with applications in communications, sensing, monitoring and intelligent signal analysis. Although deep learning, particularly convolutional neural networks (CNNs), has demonstrated strong capability for learning discriminative representations directly from raw sampled time domain signals, classification performance depends not only on network architecture but also on how the signal is represented before deep learning. Time-domain samples preserve complete temporal information but may provide limited explicit information about spectral structure, whereas Fourier- and time–frequency-based representations expose complementary characteristics that may be advantageous under noisy waveform conditions. This study extends our previous investigation of neural network-based waveform classification in additive white Gaussian noise (AWGN) by systematically evaluating the effect of signal representation using a controlled convolutional-learning framework. This study considers five classification configurations: time-domain CNN, FFT-based CNN, STFT-based CNN, wavelet-based CNN and a hybrid decision-level fusion model that combines the probabilistic outputs of the individual classifiers. To ensure a controlled comparison, the principal CNN architecture and training strategy are kept consistent across comparable representations, while only the representation and corresponding input dimensionality are changed. This study evaluates classification accuracy and confusion matrices across multiple SNR conditions. It is found that, at low SNR, and in many of the cases, the hybrid, STFT and wavelet are outperforming the time-based and FFT-based classifier, while all of them perform well at high SNRs. At an SNR of −5 dB, the accuracy performances are 78%, 62.1%, 84.35%, 82% and 90.1% for the time-domain, FFT, STFT, wavelet and the hybrid CNN classifier, respectively, for the test samples considered in this study. The hybrid deep learning that is based on the fusion of the four (time domain, FFT, STFT and wavelet) is always the best in terms of performance. When there is random amplitude, frequency and phase variations in the waveforms, the corresponding accuracy performances at SNR = −5 dB become 68.15%, 49.05%, 78.35%, 80.85% and 89.75% respectively, showing a decrease inaccuracy, but with hybrid still performing well exploiting the advantages from all representations. The results of this experiment are intended to explore the time–frequency representation-level contributions in deep-learning-based waveform classification rather than attributing performance improvements solely to increasingly complex network architectures.

Authors

Institutions

Publication Details

Journal
Signals
Published
2026-11-19
DOI
https://doi.org/10.3390/signals7060117
Primary Topic
Wireless Signal Modulation Classification
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Time–Frequency-Based CNN Waveform Classification: Comparison of Time-Domain, FFT, STFT, Wavelet and Hybrid Deep Learning

Beza Negash Getu
Signals
Wireless Signal Modulation Classification
article

Time–Frequency-Based CNN Waveform Classification: Comparison of Time-Domain, FFT, STFT, Wavelet and Hybrid Deep Learning

Beza Negash Getu
article en

Abstract

Reliable classification of noisy waveforms is an important problem in modern signal processing, with applications in communications, sensing, monitoring and intelligent signal analysis. Although deep learning, particularly convolutional neural networks (CNNs), has demonstrated strong capability for learning discriminative representations directly from raw sampled time domain signals, classification performance depends not only on network architecture but also on how the signal is represented before deep learning. Time-domain samples preserve complete temporal information but may provide limited explicit information about spectral structure, whereas Fourier- and time–frequency-based representations expose complementary characteristics that may be advantageous under noisy waveform conditions. This study extends our previous investigation of neural network-based waveform classification in additive white Gaussian noise (AWGN) by systematically evaluating the effect of signal representation using a controlled convolutional-learning framework. This study considers five classification configurations: time-domain CNN, FFT-based CNN, STFT-based CNN, wavelet-based CNN and a hybrid decision-level fusion model that combines the probabilistic outputs of the individual classifiers. To ensure a controlled comparison, the principal CNN architecture and training strategy are kept consistent across comparable representations, while only the representation and corresponding input dimensionality are changed. This study evaluates classification accuracy and confusion matrices across multiple SNR conditions. It is found that, at low SNR, and in many of the cases, the hybrid, STFT and wavelet are outperforming the time-based and FFT-based classifier, while all of them perform well at high SNRs. At an SNR of −5 dB, the accuracy performances are 78%, 62.1%, 84.35%, 82% and 90.1% for the time-domain, FFT, STFT, wavelet and the hybrid CNN classifier, respectively, for the test samples considered in this study. The hybrid deep learning that is based on the fusion of the four (time domain, FFT, STFT and wavelet) is always the best in terms of performance. When there is random amplitude, frequency and phase variations in the waveforms, the corresponding accuracy performances at SNR = −5 dB become 68.15%, 49.05%, 78.35%, 80.85% and 89.75% respectively, showing a decrease inaccuracy, but with hybrid still performing well exploiting the advantages from all representations. The results of this experiment are intended to explore the time–frequency representation-level contributions in deep-learning-based waveform classification rather than attributing performance improvements solely to increasingly complex network architectures.

SignalsVol. 7(6)
American University of Ras Al Khaimah (AE)
Wireless Signal Modulation Classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon