F2GAN: A Feature-Feedback Generative Framework for Reliable AI-Based Fault Diagnosis in Inverter-Dominated Microgrids
Swetha Rani Kasimalla, Kuchan Park · 2025
Read F2GAN: A Feature-Feedback Generative Framework for Reliable AI-Based Fault Diagnosis in Inverter-Dominated Microgrids by Swetha Rani Kasimalla, Kuchan Park (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-06677.txt.
Introduction
Swetha Rani Kasimalla and colleagues build realistic fault data for microgrids dominated by inverters, where genuine fault records are scarce and lopsided. Rare inverter failures and extreme line faults are exactly the cases a diagnosis system must catch, yet they barely appear in training data. A detailed simulated microgrid first generates a broad spread of internal and external fault scenarios.
Their generative model, F2GAN, sharpens its generated data with feedback on means, spread, links between traits, and feature matching at several levels. Classifiers trained only on the synthetic data performed strongly on real faults. A hardware-in-the-loop test bench tied to a real-time simulator reported perfect accuracy in live trials.
Catalog record
| CREATOR | Swetha Rani Kasimalla, Kuchan Park |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-06677
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
F2GAN: A Feature-Feedback Generative Framework for Reliable AI-Based Fault Diagnosis in Inverter-Dominated Microgrids Thanks: S.-R. Kasimalla, K. Park, J. Hong are with the Department of Electrical and Computer Engineering, University of Michigan – Dearborn, Dearborn, MI, 48128 USA (emails: [email protected]; [email protected]; [email protected]). Thanks: Y.-J. Kim is with the Department of Electrical Engineering, Pohang University of Science and Technology, Pohang 37673, South Korea (e-mail: [email protected]). (corresponding author: Young-Jin Kim) Swetha Rani Kasimalla, Kuchan Park Abstract Enhancing the reliability of AI-based fault diagnosis in inverter-dominated microgrids requires diverse and statistically balanced fault datasets for training. However, the intrinsic scarcity and imbalance of high-fidelity fault data, particularly for rare inverter malfunctions and extreme external line faults pose a significant limitation to dependable model training and validation. To overcome this bottleneck, this paper introduces a unified framework that first models a detailed inverter-dominated microgrid and systematically generates multiple internal and external fault scenarios to mitigate class imbalance and data scarcity. Subsequently, an enhanced generative model called F2GAN (Feature Feedback GAN) is developed to synthesize high-dimensional tabular fault data with improved realism and statistical alignment. Unlike conventional GAN variants, F2GAN integrates multi-level feedback mechanisms based on mean–variance, correlation, and feature-matching loss functions, enabling the generator to iteratively refine output distributions toward real fault feature spaces. The fidelity of the generated datasets is evaluated through quantitative and qualitative analyses. To further assess the practical utility of synthetic data, Train-on-Synthetic, Test-on-Real (TSTR) experiments are conducted, demonstrating strong generalization of downstream machine learning classifiers trained exclusively on F2GAN-generated samples. Additionally, the proposed framework is validated through a hardware-in-the-loop (HIL) fault diagnosis platform, which is integrated with a real-time simulator and a graphical user interface (GUI). This system enables real-time monitoring and classification of fault type and location using a neural network model trained on synthetic data and the real simulated data, achieving 100 % \% diagnostic accuracy under real-time testing conditions. The combined results confirm that F2GAN effectively bridges the data gap between simulated and real-world microgrid fault datasets. I Introduction The growing penetration of inverter-based distributed generators (IBDGs) within microgrids has transformed modern distribution systems into highly dynamic, inverter-dominated networks. While this transition enhances flexibility, reliability, and the integration of renewable energy sources, it simultaneously introduces several challenges in fault diagnosis. In such systems, the traditional assumption of unidirectional fault current flow no longer holds, as distributed generation alters network topology and bidirectional fault currents emerge [ 20 , 14 ] . Furthermore, inverter controllers inherently limit short-circuit currents (typically 1.5–2.0 p.u.), which complicates the sensitivity of overcurrent-based protection, especially during islanded operation. This limitation leads to rapid voltage and frequency deviations, often causing distributed units to disconnect due to out-of-zone faults before the main protection can act. Beyond external line faults, inverter-level internal faults represent another major concern, since semiconductor and converter failures account for nearly 80% of inverter outages in microgrids [ 13 ] . As inverter-based systems become increasingly critical, the demand for intelligent fault diagnosis methods that can distinguish complex fault modes has grown significantly. Machine learning (ML)-based fault classification has shown strong potential for both external and internal fault detection [ 4 , 19 , 9 , 6 ] due to its ability to learn non-linear relationships between system features. However, such models are inherently data-dependent, requiring sufficient fault event diversity and balanced class representation to achieve generalization in real operating conditions. Unfortunately, obtaining large-scale, balanced, and diverse fault datasets remains a pressing challenge. Fault experiments are limited by safety, equipment stress, and privacy constraints in real utility systems. As a result, most available data are scarce, imbalanced, and largely statistical rather than temporal, because fault diagnosis is mainly based on snapshot data. Multiple snapshots are used to train the models, which help analyze the behaviors of the faults. Therefore to alleviate data scarcity, recent studies have explored generative models such as Conditional GANs (CGAN) [ 11 ] , Wasserstein GANs (WGAN-GP) [ 22 ] , Variational Autoencoder-based GANs (VAE-GAN) [ 10 ] , and Temporal VAEs (TVAE) for various energy applications including load forecasting [ 12 ] , renewable scenario generation [ 18 ] , and transient stability analysis [ 3 ] . For instance, an instability sample generation framework using GANs was proposed in [ 3 ] for online transient stability assessment (TSA), where a GAN-based model was trained on temporal dynamic trajectories to balance the long-tailed distribution of instability cases. Similarly, an unbalanced graph-GAN architecture (UG-GAN) was developed in [ 16 ] to synthetically generate three-phase unbalanced distribution system connectivity, capturing structural relationships through random-walk representations of network graphs rather than numeric fault data. Both works demonstrate the promise of GANs in generating rare or unobservable operating scenarios; however, their focus remains on system-level or graph-based temporal data, rather than physical feature-level fault patterns. In the renewable energy domain, several studies have employed GAN-based architectures for time-series prediction and scenario generation. For example, [ 17 ] utilized a time-series GAN for wind power output correction under extreme weather conditions, while [ 18 ] introduced a style-based scenario generator coupled with a sequence encoder to reproduce spatial–temporal renewable dynamics. These approaches effectively capture temporal dependencies but lack mechanisms to ensure per-feature statistical coherence, which is crucial when dealing with voltage–current fault signatures. A transfer-learning-integrated improved GAN was proposed in [ 7 ] to augment operation data under data-scarce conditions, demonstrating improved generalization with limited samples, yet without validation under dynamic or real-time conditions. More recently, diffusion-based generative models such as EnergyDiff [ 8 ] have been applied to high-resolution time-series generation in energy systems, emphasizing temporal consistency but still neglecting multi-feature physical dependencies. Generative approaches have also been applied to extreme event modeling and fault data augmentation. An auxiliary classifier GAN (ACGAN) was presented in [ 11 ] to address data imbalance in fault detection and classification tasks, with a focus on improving the representation of minority classes. Likewise, the improved WGAN-GP model in [ 22 ] generated synthetic samples for partial-discharge source classification, enabling stable training and improved performance on small datasets. While both works demonstrate the effectiveness of GANs for mitigating data imbalance, they primarily target single-domain classification tasks and do not incorporate statistical feedback mechanisms to preserve inter-feature relationships. Similarly, [ 5 ] proposed MultiLoad-GAN to generate realistic load profiles for different consumers, and [ 12 ] presented a super-resolution GAN (ProfileSR-GAN) for upsampling low-resolution load profiles, both confirming the utility of GANs for time-series load data. However, these approaches are limited to forecasting or resolution enhancement rather than diagnostic data synthesis. A hybrid VAE-GAN architecture was introduced in [ 10 ] to generate energy consumption time-series in smart homes, achieving diverse yet coherent load patterns. Although such hybrid models bridge generative diversity with distribution learning, they are inherently designed for continuous temporal data, rather than high-dimensional tabular datasets like fault snapshots in microgrids. While these prior works leverage GANs for time-series augmentation, network topology synthesis, or renewable forecasting, they predominantly operate in offline, statistical, or temporal domains and need to be tested on snapshot datasets. Crucially, not all GAN architectures are universally applicable across all types of data. The structure and design of a GAN must be tailored to the specific characteristics of the dataset and