Darong Liang
| Jian Yu* | Hui Wang | Xiaodan Chen | Chengxuan Huang | Ze Li | Yuanyuan Luo | Xiangqing Wei
© 2025 The authors. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
Generative Adversarial Networks (GANs) have demonstrated exceptional capabilities in signal processing tasks like noise reduction and feature enhancement, yet traditional implementations face limitations including mode collapse and restricted output diversity. To address these challenges, this study proposes the Entropy-Maximized Generative Adversarial Network (EM-GAN), a novel framework for signal denoising and feature enhancement. By integrating entropy maximization into adversarial training, EM-GAN enhances signal diversity, mitigates mode collapse, and improves training stability. The framework employs entropy-driven activation functions and loss functions to suppress noise while preserving critical signal features. We introduce an entropy-regularized objective function that incentivizes the generator to produce high-entropy outputs, enabling more comprehensive modeling of underlying signal distributions. Theoretical and experimental analyses confirm the model's stability improvements and superior performance over conventional GAN variants in output quality and diversity. These results highlight EM-GAN's potential for robust noise suppression and feature preservation in modern signal processing applications, representing a significant methodological advancement in generative model design.
signal denoising, feature enhancement, entropy maximization, Generative Adversarial Networks, signal processing
Signal denoising and feature enhancement are fundamental challenges in signal processing, with broad implications for applications such as image restoration, audio enhancement, and Intelligent Logistics signal analysis. The core problem is how to recover clean, high-fidelity signals from noisy or degraded inputs while simultaneously preserving and enhancing important features in the data. Traditional filtering and reconstruction techniques often struggle to remove noise without blurring critical signal details. Recent advances in deep learning, particularly GANs, have shown promise in learning to produce high-quality signal reconstructions. However, standard GAN models suffer from limitations like mode collapse – where the generator produces overly narrow or average outputs – and training instability, which together hinder their effectiveness in denoising tasks that require capturing diverse signal characteristics.
This paper addresses the above challenge by proposing a novel approach to improve signal denoising and feature preservation using generative modeling. We introduce the EM-GAN, a new GAN framework that explicitly integrates an entropy maximization principle to overcome the diversity and stability issues of conventional GANs. Inspired by the thermodynamic principle that systems evolve toward higher entropy (greater disorder and uncertainty), EM-GAN incorporates an entropy-based regularization into the generator’s training. By maximizing the entropy of the generator’s output distribution, the model is encouraged to explore a broader range of plausible outputs for a given input. This leads to richer and more diverse signal representations, ensuring that fine-grained features are not averaged out. In effect, EM-GAN produces denoised signals that retain important details, and it achieves more stable training convergence. The entropy-driven objective provides a strong regularization that mitigates mode collapse and delivers more consistent optimization, resulting in enhanced signal quality and reliability.
The main contributions of this work are as follows:
In summary, EM-GAN provides a significant methodological advancement for signal processing. By integrating the principle of entropy maximization into GAN training, it offers a powerful solution for generating clean, high-quality signals with enriched features. This work opens up new possibilities for applying generative models to critical signal processing tasks – from image and audio denoising to data augmentation and beyond – where maintaining signal integrity and diversity is paramount. The proposed framework not only addresses a key limitation in GAN-based signal reconstruction but also broadens the scope of GAN applications in signal processing by combining thermodynamic insights with deep learning for enhanced signal processing performance.
2.1 Entropy maximization in GANs for signal processing
Entropy-based optimization has been increasingly incorporated into GAN training to enhance output diversity and training stability, which are critical for signal processing tasks. Traditional GANs frequently suffer from mode collapse, where the generator fails to produce diverse outputs, leading to limited applicability in tasks requiring variability in generated signals. To address this, researchers have proposed entropy-regularized approaches to ensure richer feature generation. For instance, Manifold-Preserving GAN (MP-GAN) [1] integrates entropy maximization on the latent space distribution to maintain structural diversity in generated data, effectively mitigating mode collapse and enhancing learning stability. InfoMax-GAN [2], a model designed to maximize mutual information, indirectly promotes entropy maximization by ensuring that each generated sample retains distinct information characteristics, leading to more expressive and robust representations. Furthermore, Variational Entropy Regularization (VER-GAN) [3] directly optimizes the generator’s entropy to encourage output diversity, significantly improving stability in training and ensuring a more comprehensive reconstruction of signals.
While these entropy-based methods have been explored in the context of image synthesis and classification, their adoption in signal processing remains limited. Signal processing applications, particularly denoising and feature enhancement, require models capable of preserving fine-grained signal features while suppressing noise-induced distortions. The integration of entropy maximization in GANs offers a promising approach to solving these issues by enhancing diversity, improving feature retention, and stabilizing adversarial training. These motivations drive the development of the EM-GAN [4], which explicitly integrates entropy regularization into GAN training for signal denoising and enhancement.
2.2 Traditional GAN-based signal denoising methods
GANs have demonstrated strong potential in signal denoising by learning to map noisy inputs to clean outputs. Several studies have leveraged adversarial training to suppress noise in signals, including applications in EEG signal restoration, biomedical waveform processing, and sensor noise removal. For example, WGAN-based models have been employed for EEG artifact removal, showing promising results in extracting meaningful brainwave signals while filtering out physiological distortions [5]. Asymmetric GANs have also been developed for EEG denoising, allowing models to learn from unpaired noisy and clean data, avoiding explicit noise label dependency.
Despite their effectiveness, traditional GAN-based denoisers face persistent limitations:
To address these shortcomings, EM-GAN introduces entropy maximization as a key component of the training process. Unlike traditional GANs that optimize purely for adversarial loss, EM-GAN incorporates an entropy-based regularization term, which:
These contributions position EM-GAN as a significant improvement over existing GAN-based denoising approaches, offering a more stable, diverse, and robust solution for signal enhancement in real-time and high-fidelity applications.
2.3 Real-time signal analysis
Deep learning has made significant advancements in real-time industrial signal analysis, playing a crucial role in applications such as Intelligent Logistics Management Systems (ILMS) by enabling the efficient processing of sensor data. In modern industrial settings, vast amounts of real-time signals—such as GPS tracking data, temperature fluctuations in storage environments, and vibration measurements from vehicles and equipment—are continuously generated. The effective analysis of these signals is essential for enhancing tracking accuracy, optimizing condition monitoring, and improving predictive maintenance, ultimately leading to increased operational efficiency and system reliability.
2.3.1 GPS signal analysis
Recurrent neural networks (RNNs) [13] and their variants, such as LSTMs and Bi-LSTMs, have been widely applied to GPS trajectory modeling for route prediction and anomaly detection. By learning sequential dependencies in vehicle movement data, these models can forecast future routes and identify deviations from expected paths. Studies have demonstrated that LSTM-based models achieve over 96% accuracy in vehicle destination prediction by leveraging historical trajectory data [14]. These approaches have significantly enhanced industrial logistics efficiency by improving route optimization and anomaly detection in fleet management systems.
2.3.2 Temperature monitoring in logistics
Maintaining stable temperature conditions is crucial in logistics, particularly for cold-chain management in pharmaceutical and food industries. Deep learning techniques, such as autoencoders, have been employed to detect temperature anomalies in real time. Convolutional autoencoders have shown high effectiveness in reconstructing expected temperature profiles and flagging deviations, achieving over 92% fault detection accuracy in cold storage monitoring applications [15]. Such models help prevent product spoilage by enabling early detection of refrigeration failures or sensor faults.
2.3.3 Vibration-based predictive maintenance
Sensors embedded in industrial logistics vehicles and equipment continuously capture vibration data, which can be analyzed to predict mechanical failures. Hybrid CNN–LSTM architectures have been utilized for time-series vibration analysis, learning both spatial and temporal features indicative of system health. Research indicates that combining convolutional feature extraction with sequential modeling significantly improves fault detection accuracy, making predictive maintenance in industrial systems more reliable [16]. Such models enable early identification of mechanical degradation, allowing industrial logistics companies to reduce downtime and maintenance costs.
2.4 Summary
This section highlights the significance of entropy maximization in GANs, its role in addressing the limitations of traditional GAN-based denoising methods, and the expanding application of deep learning in intelligent industrial sensor analysis. EM-GAN builds on these advancements by incorporating entropy-driven optimization to enhance signal diversity, training stability, and feature retention, making it a robust solution for real-time signal enhancement in industrial applications and beyond.
While recent studies have begun to apply entropy-based regularization in signal processing tasks—such as ECG denoising [17], speech enhancement [18], and wavelet-domain physiological signal filtering [19]—these efforts remain highly task-specific and rarely incorporate adversarial or generative mechanisms. Moreover, their applicability across diverse signal types and noise conditions is limited. This highlights a critical gap in current research and underscores the need for a unified generative framework with entropy-aware mechanisms, motivating the design of EM-GAN.
In contrast to existing entropy-regularized GANs such as InfoMax-GAN and VER-GAN—which primarily target image generation or representation learning—EM-GAN is specifically designed for signal denoising tasks. It introduces entropy regularization not only at the loss-function level but also within the network architecture via a novel Entropy-Sensitive Activation (ESA) mechanism. By dynamically adjusting activations according to mini-batch entropy estimates, ESA enables EM-GAN to effectively preserve transient and fine-grained signal features, such as spikes and abrupt changes, while maintaining strong noise suppression. This dual integration of entropy principles makes EM-GAN particularly suitable for real-time signal processing scenarios where structural fidelity is critical.
The EM-GAN proposes an entropy-driven framework for signal denoising and feature enhancement, addressing key limitations of traditional GAN-based denoisers, such as mode collapse and training instability. By embedding entropy maximization into the loss function and network architecture, EM-GAN ensures diverse, high-fidelity signal reconstructions with improved feature preservation.
This section outlines:
(1): The theoretical basis and mathematical formulation of entropy maximization within GAN training.
(2): The network architecture and processing pipeline optimized for retaining critical signal features.
(3): The training strategy and stabilization mechanisms that underpin robust and consistent performance.
3.1 Entropy-maximized GAN architecture
Traditional GAN-based denoisers often encounter mode collapse, where the generator produces a limited range of outputs, resulting in the loss of fine signal details during reconstruction. Additionally, training instability can lead to divergence or degraded signal quality. To address these issues, the proposed EM-GAN incorporates entropy regularization into the training process. This approach enhances output diversity to mitigate mode collapse, stabilizes adversarial training dynamics, and improves the preservation of signal variations critical for high-fidelity denoising.
3.1.1 Mathematical formulation
To estimate the mini-batch entropy ${{H}_{batch}}\left( x \right)$, we apply a kernel density estimation (KDE) method with Gaussian kernels in the generator’s feature space. This provides an efficient approximation of entropy during training without directly operating in the high-dimensional signal domain. Although KDE may introduce bias due to bandwidth sensitivity and sample sparsity, we mitigate these effects by using adaptive bandwidth selection (via Silverman's rule) and averaging entropy across multiple mini-batches. This ensures that the entropy signal remains stable and meaningful for guiding both the entropy-regularized loss and the ESA. More advanced estimators such as k-nearest neighbor or plug-in entropy methods may be explored in future work.
EM-GAN integrates entropy maximization in two key components:
Together, these mechanisms help the generator produce denoised outputs that retain fine-grained details while effectively suppressing noise.
The objective of a standard GAN is defined within a minimax framework:
$\underset{G}{\mathop{\min }}\,\underset{D}{\mathop{\max }}\,{{\mathbb{E}}_{x\sim{{p}_{data}}}}\left[ \log D\left( x \right) \right]+{{\mathbb{E}}_{z\sim{p}\left( z \right)}}\left[ \log \left( 1-D\left( G\left( z \right) \right) \right) \right]$
To promote diversity in the generator’s outputs, EM-GAN augments this objective with an entropy regularization term, modifying the generator’s loss as follows:
${{L}_{G}}={{L}_{{{G}_{adv}}}}-\lambda H\left( {{p}_{g}} \right)$
where,
By maximizing $H\left( {{p}_{g}} \right)$, EM-GAN encourages the generator to produce a broader range of plausible clean signals, preventing over-smoothing or repetitive outputs commonly observed in traditional GANs.
3.1.2 ESA function
In addition to the loss function, EM-GAN integrates entropy awareness into the network architecture through a novel ESA function, defined as:
$ESAH\left( x \right)=x·exp\left( \alpha {{H}_{\text{batch}}}\left( x \right) \right)$
where,
The ESA function dynamically adjusts activations based on entropy levels: higher entropy amplifies activations to preserve signal diversity and prevent excessive smoothing, while lower entropy attenuates activations to ensure effective noise suppression. This adaptive mechanism enables EM-GAN to balance feature retention and denoising efficacy, particularly excelling in preserving subtle signal details essential for signal processing tasks.
Unlike attention-based adaptive mechanisms that reweight features based on learned correlations, ESA modulates activation strength using entropy statistics. This design is specifically suited to preserving transient details in 1D signal denoising, without introducing additional learnable parameters.
3.2 Entropy-driven signal denoising and feature preservation
Effective signal denoising requires removing unwanted noise while preserving essential signal features such as frequency components, transient structures, and amplitude variations. Traditional filtering-based denoising methods, such as Wiener filtering and wavelet transforms, often introduce distortions or fail to adapt to varying noise distributions. Similarly, conventional GAN-based denoisers suffer from mode collapse, where the generator produces oversmoothed and repetitive outputs, leading to a loss of fine-grained signal details.
To address these limitations, EM-GAN introduces an entropy-regularized learning framework, dynamically balancing noise suppression and feature retention. By maximizing entropy, the generator avoids collapsing to a limited set of solutions and instead learns a diverse set of noise-free reconstructions, ensuring that the intrinsic characteristics of the signal are preserved.
3.2.1 Algorithmic approach to entropy-driven signal denoising
The fundamental challenge in signal denoising is ensuring that noise is effectively removed without degrading signal integrity. To achieve this, EM-GAN employs:
The modified generator objective function is formulated as:
${{L}_{G}}=L_{G}^{\text{adv}}-\lambda H\left( {{p}_{g}} \right)+\gamma ||G\left( {{x}_{\text{noisy}}} \right)-{{x}_{\text{clean}}}||_{1}$
where,
To control the trade-off between adversarial learning and entropy regularization, we define λ as a dynamic parameter that increases over training epochs. Specifically, we use a simple exponential schedule:
$\lambda \left( t \right)=\lambda_0 \times \left( 1-{{e}^{-\alpha t}} \right)$
where, $\lambda _0$ is the maximum weight (set to 1.0), α is a growth rate constant (typically 0.05), and t denotes the current epoch. This design allows entropy regularization to gradually take effect as training stabilizes, avoiding early training instability while enhancing diversity in later stages.
This formulation ensures that noise suppression does not lead to over-smoothing, preserving the temporal and spectral characteristics of the signal.
3.2.2 Processing pipeline for EM-GAN signal denoising
To effectively enhance and restore noisy signals, EM-GAN follows a structured four-step processing pipeline:
Step 1: Preprocessing and Data Augmentation
Time-domain transformations (random shifts, amplitude modulation).
Frequency-domain augmentations (Fourier filtering, spectral distortions).
Artificial noise addition (Gaussian, Poisson, impulse noise) to improve robustness.
Step 2: Feature-Aware Signal Generation
The generator $G$ in EM-GAN is designed to enhance noisy inputs while retaining important features:
Generator Processing Flow:
Step 3: Entropy-Driven Discriminator Evaluation
The discriminator $D$ functions as an adaptive evaluator, ensuring that the denoised output maintains realism while removing noise. Unlike conventional discriminators, EM-GAN’s $D$ integrates:
Discriminator Processing Flow:
Step 4: Entropy-Optimized Training Update
To prevent instability in GAN training, EM-GAN employs:
This adaptive training process ensures that EM-GAN achieves robust convergence while maintaining high-fidelity reconstructions.
3.2.3 Visualizing EM-GAN signal denoising performance
Figure 1 illustrates the denoising performance of EM-GAN, highlighting how the model reconstructs a clean signal from noisy input while retaining key features.
Figure 1. EM-GAN signal denoising process
Key Observations from Figure 1:
This visualization confirms that EM-GAN’s entropy-driven approach enables high-fidelity signal restoration, outperforming both traditional filtering methods and standard GAN-based denoisers.
3.2.4 Comparative analysis: EM-GAN vs. traditional denoising methods
A direct comparison between EM-GAN and other denoising techniques is presented in Table 1, showcasing its superior performance in feature preservation, noise suppression, and training stability.
Table 1. Comparison between EM-GAN and other denoising techniques
|
Method |
Noise Reduction |
Feature Preservation |
Output Diversity |
Training Stability |
|
Wiener Filter |
Moderate |
Loss of fine details |
Limited |
Stable |
|
Wavelet Denoising |
Moderate |
Some feature retention |
Limited |
Stable |
|
Vanilla GAN |
High |
Often over-smooths |
Mode collapse |
Unstable |
|
EM-GAN |
High |
Strong feature retention |
High diversity |
Robust training |
To facilitate a clearer understanding of the architectural distinctions between EM-GAN and related entropy-regularized GANs (e.g., InfoMax-GAN, VER-GAN), we have included a detailed comparison in Table 2, outlining differences in entropy regulation mechanisms, activation strategies, and signal denoising capabilities.
Table 2. Architectural comparison between EM-GAN and other entropy-regularized GAN models
|
Model |
Entropy Regularization Location |
Activation Function Type |
Target Application |
Feature Preservation Strategy |
Mode Collapse Mitigation |
|
InfoMax-GAN |
Latent space (via mutual information) |
Standard ReLU / Leaky ReLU |
Image generation, classification |
Information maximization in latent code |
Partial (via mutual info) |
|
VER-GAN |
Generator loss (KL divergence) |
Standard activations |
Generic generation tasks |
Encourages sample diversity through entropy loss |
Yes |
|
EM-GAN |
Generator loss + activation layer |
ESA |
Signal denoising, feature enhancement |
Mini-batch entropy-driven activation to retain transient details |
Yes (adaptive entropy feedback). |
3.2.5 Summary: How EM-GAN enhances signal denoising performance
The EM-GAN employs entropy-regularized learning to achieve superior signal denoising while preserving structural integrity, effectively overcoming challenges such as mode collapse, feature loss, and training instability inherent in traditional GAN-based methods. The key mechanisms contributing to its performance are outlined below:
By integrating entropy maximization into the adversarial learning framework, EM-GAN achieves high-fidelity signal denoising that surpasses traditional filtering methods and conventional GAN-based approaches in both noise suppression and feature retention. This synergy of entropy-driven regularization and architectural innovations positions EM-GAN as a robust and effective solution for real-world signal processing applications, with its efficacy further validated in subsequent experimental evaluations.
This section presents a comprehensive evaluation of EM-GAN’s signal denoising performance, comparing it against traditional filtering techniques and existing GAN-based denoising models. The experiments are conducted across multiple benchmark datasets, and the results are analyzed based on quantitative performance metrics, visual comparisons, computational efficiency, and an ablation study.
4.1 Experimental setup
4.1.1 Datasets
The performance of EM-GAN is evaluated using multiple benchmark datasets containing diverse signal types with varying noise levels:
Each dataset includes both clean reference signals and their corresponding noisy versions, allowing precise evaluation of denoising quality.
4.1.2 Baseline methods
The performance of EM-GAN is compared against several widely used denoising methods, including traditional filtering techniques and deep learning models (Table 3).
Table 3. Comparison between EM-GAN and other denoising methods
|
Method |
Type |
Description |
|
Wiener Filter |
Traditional Filtering |
Assumes stationary noise; effective at low SNR but struggles with nonstationary signals. |
|
Wavelet Denoising |
Transform-Based Filtering |
Decomposes signals into frequency components; effective for structured noise but may lead to feature loss. |
|
Denoising Autoencoder (DAE) |
Deep Learning |
Learns to reconstruct clean signals from noisy inputs but lacks adversarial training for high-fidelity restoration. |
|
Vanilla GAN |
GAN-Based Denoising |
Uses adversarial learning but often exhibits mode collapse, leading to oversmoothed outputs. |
|
EM-GAN (Proposed) |
Entropy-Regularized GAN |
Incorporates entropy maximization to enhance diversity, stability, and feature retention. |
4.1.3 Evaluation metrics
The performance of each denoising method is assessed using the following standard metrics:
While formal significance tests were not included, all reported metrics were averaged across multiple trials with fixed seeds to ensure consistency. Future studies will incorporate statistical testing to further validate performance differences.
4.2 Quantitative performance evaluation
4.2.1 Comparative analysis of denoising performance
The quantitative results for each method are summarized in Figure 2, which presents the average performance across all datasets.
Figure 2. Comparative performance analysis of denoising methods
(SNR, PSNR, SSIM, and MSE comparison of different denoising methods.)
The results demonstrate that EM-GAN consistently achieves the highest SNR, PSNR, and SSIM scores while maintaining a low MSE, indicating improved denoising performance.
4.2.2 Visual comparison of denoised signals
A visual comparison of denoised outputs is provided in Figure 3.
Figure 3. Visual comparison of signal denoising methods
(This figure presents a side-by-side comparison of noisy signals, clean ground-truth signals, and denoised outputs obtained from different methods)
The visual analysis highlights that traditional filtering methods effectively suppress noise but introduce noticeable distortions. Deep learning-based methods such as DAE and Vanilla GAN provide improved denoising but tend to oversmooth the signals. The proposed EM-GAN method achieves superior noise reduction while preserving fine details, closely resembling the clean ground truth.
4.3 Ablation study: Effect of entropy maximization
An ablation study is conducted to evaluate the impact of entropy regularization in EM-GAN. The performance of the full model is compared to a variant without entropy loss, as illustrated in Figure 4.
Figure 4. Impact of entropy regularization on denoising performance
The results indicate that entropy maximization plays a crucial role in improving denoising performance. Without entropy regularization, the GAN-based model exhibits mode collapse and reduced structural fidelity, leading to increased distortion in the reconstructed signals.
4.4 Computational efficiency and real-time feasibility
The evaluation is conducted on a system with the following hardware specifications:
To provide a meaningful analysis, EM-GAN’s inference time is compared against traditional filtering techniques and deep learning-based denoisers. The results are summarized in Figure 5, where inference time is plotted against signal length for different methods
Figure 5. Computational efficiency of EM-GAN
(Computational Efficiency Comparison of Denoising Methods, comparing inference times for short (100ms) and long (5000ms) signals across different denoising techniques.)
Key Observations:
This figure reinforces that EM-GAN is computationally feasible for real-time applications while outperforming other GAN-based models in terms of scalability and stability.
4.5 Discussion
4.5.1 Comparison with baseline methods
The experimental findings indicate that EM-GAN outperforms traditional filtering and deep learning-based denoising techniques in terms of both quantitative metrics and qualitative assessments. The high SNR and PSNR scores reflect the model’s ability to suppress noise while preserving critical signal features. The superior SSIM and lower MSE values further confirm the fidelity of the reconstructed signals.
4.5.2 Role of entropy maximization
The ablation study demonstrates that entropy regularization is a key factor in enhancing the performance of EM-GAN. Without this component, the model struggles with mode collapse, resulting in less diverse and lower-quality outputs. Entropy maximization ensures that the generator produces feature-rich, high-fidelity reconstructions, preserving both high- and low-frequency signal components.
4.5.3 Practical implications and real-world applications
The computational efficiency analysis suggests that EM-GAN is capable of real-time operation, making it applicable in various domains, including:
While our current real-time performance evaluation is conducted on high-performance hardware, the lightweight architecture of EM-GAN—featuring a shallow convolutional backbone and non-parametric entropy-based components—indicates strong potential for deployment on edge or embedded platforms. With further optimization techniques such as model pruning, quantization, or hardware-specific acceleration, EM-GAN can be adapted to resource-constrained environments. Future work will explore these deployment pathways to extend its practical applicability.
4.5.4 Limitations and future work
While EM-GAN demonstrates state-of-the-art performance in signal denoising, several limitations should be acknowledged. First, its generalization to unseen or complex noise types has not been fully evaluated. For instance, EM-GAN may experience performance degradation when exposed to non-Gaussian noise distributions, such as heavy-tailed or temporally correlated interference, which are underrepresented in the current training datasets. This reflects the model’s sensitivity to the statistical characteristics of the input noise.
Moreover, the current framework has not been tested on adversarial or non-stationary noise scenarios, including perturbations designed to exploit model vulnerabilities or environments exhibiting dynamic distribution shifts. These factors may pose significant challenges to the model's robustness in practical deployment. To mitigate these issues, future work will investigate domain adaptation techniques, such as adversarial adaptation, self-supervised fine-tuning using target-domain data, and meta-learning strategies for rapid adjustment to unfamiliar noise patterns.
Second, the training process of EM-GAN is computationally demanding, which may limit its deployment in resource-constrained environments. Optimization strategies, such as model compression or hardware-aware acceleration, will be essential for enabling real-time performance on edge devices.
Finally, the current design does not explicitly incorporate domain-specific priors or hybrid architectural elements. Future research could explore combining EM-GAN with adaptive filtering methods or physics-informed components to enhance robustness. Expanding the diversity of noise types during training will also be considered to improve generalization across a broader range of real-world scenarios. Recent developments in diffusion-based models [20, 21] have also shown promise in signal generation tasks, although their inference efficiency remains a challenge for real-time denoising applications.
4.6 Summary of experimental results
These results establish EM-GAN as an effective, stable, and computationally efficient solution for signal denoising, with promising applicability across multiple domains. Compared to both traditional filtering techniques and recent GAN-based models, EM-GAN consistently demonstrates superior performance in noise suppression and feature preservation. Entropy regularization emerges as a critical component in maintaining structural fidelity and mitigating mode collapse. Moreover, the model exhibits competitive inference efficiency, reinforcing its suitability for real-time deployment in biomedical signal processing, speech enhancement, and structural health monitoring scenarios.
This work introduced EM-GAN, an entropy-maximized GAN designed for signal denoising and feature preservation. By integrating entropy regularization into adversarial learning, EM-GAN effectively mitigates mode collapse, enhances signal diversity, and preserves essential structural features while suppressing noise. Future work will focus on optimizing computational efficiency, extending the model to diverse noise environments, and exploring hybrid approaches that combine EM-GAN with adaptive filtering techniques to further enhance robustness and generalization. The findings demonstrate that EM-GAN is a high-performing, stable, and computationally efficient solution for real-time signal denoising, offering promising advancements in modern signal processing applications.
This paper was supported by the Foundation of Improving Academic Ability in University for Young Scholars of Guangxi (Grants No.: 2024KY1070; 2025KY1357); the Major Research Project of Liuzhou Polytechnic University (Grants No.: 2023KA06;2023SA02; 2024KA03; 2024KA09).
Darong Liang, Jian Yu, and Hui Wang contributed equally to this work. Darong Liang led the experimental design and conducted model development. Jian Yu proposed the original concept of entropy-driven generative modeling, supervised the overall project, and coordinated team collaboration. Hui Wang contributed to the mathematical formulation and implementation of entropy-regularized training strategies. Xiaodan Chen assisted in industrial application analysis and data preprocessing. Chengxuan Huang and Ze Li supported benchmarking and visualization. Yuanyuan Luo and Xiangqing Wei contributed to the literature review and manuscript editing. All authors read and approved the final manuscript.
[1] Zhou, Y., Wang, J., Tang, J., Gou, C., Jiang, Z., Li, D., Ng, S.K. (2023). MP-GAN: Cyber-attack detection and localization for cyber-physical systems with multi-process generative adversarial networks. In 2023 International Conference on Artificial Intelligence of Things and Systems (AIoTSys), Xi'an, China, pp. 186-193. https://doi.org/10.1109/AIoTSys58602.2023.00049
[2] Lee, K.S., Tran, N.T., Cheung, N.M. (2021). InfoMax-GAN: Improved adversarial image generation via information maximization and contrastive learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, pp. 3942-3952. https://doi.org/10.1109/WACV48630.2021.00399
[3] Lambert, A., Boots, B. (2021). Entropy regularized motion planning via stein variational inference. arXiv preprint arXiv:2107.05146. https://doi.org/10.48550/arXiv.2107.05146
[4] Yu, J., Wang, H., Huang, C., Li, Z. (2024). Entropy-Maximized Generative Adversarial Network (EM-GAN) based on the thermodynamic principle of entropy increase. Traitement du Signal, 41(6): 3255-3264. https://doi.org/10.18280/ts.410641
[5] Gu, S., Yang, L., Du, Y., Chen, G., Walter, F., Wang, J., Knoll, A. (2022). A review of safe reinforcement learning: Methods, theory and applications. arXiv preprint arXiv:2205.10330. https://doi.org/10.48550/arXiv.2205.10330
[6] Dai, Z., Zhao, L., Wang, K., Zhou, Y. (2024). Mode standardization: A practical countermeasure against mode collapse of GAN-based signal synthesis. Applied Soft Computing, 150: 111089. https://doi.org/10.1016/j.asoc.2023.111089
[7] Xu, K., Liao, L., Xiao, J., Chen, C., Wu, H., Yan, Q., Lin, W. (2024). Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, pp. 2662-2672. https://doi.org/10.1109/CVPR52733.2024.00257
[8] Guo, Y., Cheng, T., Yang, Z., Huang, Y., Li, M., Wang, T. (2024). A systematic review and meta-analysis of balance training in patients with chronic ankle instability. Systematic Reviews, 13(1): 64. https://doi.org/10.1186/s13643-024-02455-x
[9] Liu, D. (2024). The effects of segmentation on cognitive load, vocabulary learning and retention, and reading comprehension in a multimedia learning environment. BMC Psychology, 12(1): 4. https://doi.org/10.1186/s40359-023-01489-5
[10] Welfert, M., Kurri, G.R., Otstot, K., Sankar, L. (2024). Addressing GAN training instabilities via tunable classification losses. IEEE Journal on Selected Areas in Information Theory, 5: 534-553. https://doi.org/10.1109/JSAIT.2024.3415670
[11] Hu, S., Gao, F., Zhou, X., Dong, J., Du, Q. (2024). Hybrid convolutional and attention network for hyperspectral image denoising. IEEE Geoscience and Remote Sensing Letters, 21: 5504005. https://doi.org/10.1109/LGRS.2024.3370299
[12] Li, X., Zheng, Y., Zhang, J., Dang, S., Nallanathan, A., Mumtaz, S. (2024). Finite SNR diversity-multiplexing trade-off in hybrid ABCom/RCom-assisted NOMA networks. IEEE Transactions on Mobile Computing, 23(10): 9108-9119. https://doi.org/10.1109/TMC.2024.3357753
[13] Mienye, I.D., Swart, T.G., Obaido, G. (2024). Recurrent neural networks: A comprehensive review of architectures, variants, and applications. Information, 15(9): 517. https://doi.org/10.3390/info15090517
[14] Sayegh, H.R., Dong, W., Al-madani, A.M. (2024). Enhanced intrusion detection with LSTM-based model, feature selection, and SMOTE for imbalanced data. Applied Sciences, 14(2): 479. https://doi.org/10.3390/app14020479
[15] Madani, S.S., Ziebert, C., Vahdatkhah, P., Sadrnezhaad, S.K. (2024). Recent progress of deep learning methods for health monitoring of lithium-ion batteries. Batteries, 10(6): 204. https://doi.org/10.3390/batteries10060204
[16] Lin, K., Pan, J., Xi, Y., Wang, Z., Jiang, J. (2024). Vibration anomaly detection of wind turbine based on temporal convolutional network and support vector data description. Engineering Structures, 306: 117848. https://doi.org/10.1016/j.engstruct.2024.117848
[17] Parbat, D., Chakraborty, M. (2024). Multiscale entropy analysis of single lead ECG and ECG derived respiration for AI based prediction of sleep apnea events. Biomedical Signal Processing and Control, 87: 105444. https://doi.org/10.1016/j.bspc.2023.105444
[18] Wang, L., Liu, Y., Yu, Z., Gao, S., Mao, C., Huang, Y., Dong, L. (2025). SECodec: Structural entropy-based compressive speech representation codec for speech language models. In Proceedings of the AAAI Conference on Artificial Intelligence, Washington, DC, USA, pp. 25380-25388. https://doi.org/10.1609/aaai.v39i24.34726
[19] Ananthi, A., Subathra, M.S.P., George, S.T., Sairamya, N.J. (2024). Entropy-based feature extraction for classification of EEG signal using Lifting Wavelet Transform. Przegląd Elektrotechniczny, 100: 146-150. https://doi.org/10.15199/48.2024.09.27
[20] Yi, H., Hou, L., Jin, Y., Saeed, N.A., Kandil, A., Duan, H. (2024). Time series diffusion method: A denoising diffusion probabilistic model for vibration signal generation. Mechanical Systems and Signal Processing, 216: 111481. https://doi.org/10.1016/j.ymssp.2024.111481
[21] Lemercier, J.M., Richter, J., Welker, S., Moliner, E., Välimäki, V., Gerkmann, T. (2025). Diffusion models for audio restoration: A review. IEEE Signal Processing Magazine, 41(6): 72-84. https://doi.org/10.1109/MSP.2024.3445871
