Scipediacontent: Scipediacontent moved page Review 538456560717 to Irfan et al 2025a

2025-12-19T09:18:14Z

Scipediacontent moved page Review 538456560717 to Irfan et al 2025a

JSanchez: JSanchez moved page Draft Sanchez Pinedo 521166662 to Review 538456560717

2025-12-19T08:50:37Z

JSanchez moved page Draft Sanchez Pinedo 521166662 to Review 538456560717

JSanchez: Created page with " == Abstract ==
Wind turbine reliability is critical for sustainable energy production, yet fault diagnosis faces challenges due to data privacy concerns, heterogeneous op..."

2025-12-19T08:50:35Z

Created page with " == Abstract == <p>Wind turbine reliability is critical for sustainable energy production, yet fault diagnosis faces challenges due to data privacy concerns, heterogeneous op..."

New page

== Abstract ==

<p>Wind turbine reliability is critical for sustainable energy production, yet fault diagnosis faces challenges due to data privacy concerns, heterogeneous operational conditions, and resource constraints in distributed wind farms. Traditional centralized Machine Learning (ML) approaches struggle with these issues, necessitating decentralized solutions. This study introduces the Adaptive Federated Fault Diagnosis (AF2D) framework, a novel Federated Learning (FL) approach for wind turbine fault diagnosis that ensures data privacy while addressing non-i.i.d. data distributions. Using a dataset of 35 uniaxial vibration recordings from six turbines at the University of Mustansiriyah, AF2D leverages two key modules: Adaptive Model Aggregation (AMA) and Lightweight Model Optimization (LMO). AMA employs Jensen-Shannon divergence and cosine similarity to adaptively aggregate local model updates, mitigating data heterogeneity, while LMO applies structured pruning (60% filter reduction) and 8bit quantization to enable deployment on resource-constrained SCADA systems. Results show AF2D achieves 91.3% accuracy (±1.2%, 95% confidence interval), a 3.5% improvement over FedAvg (87.8%± 1.4%), with statistical significance (p < 0.05), and outperforms state-of-the-art methods like Clustered FL (88.5%) and Privacy-Preserving FL (87.2%). LMO reduces inference time by 64.44% and memory usage by 53.71%, enhancing edge deployment feasibility. However, the small dataset raises overfitting risks, and scalability tests reveal a threefold communication cost increase (54.5 to 150.6 MB) for 18 clients, mitigated by proposed compression (30%–50% reduction) and asynchronous updates (20%–40% overhead reduction). Privacy is maintained with a differential privacy guarantee of= 1.0, though advanced techniques like secure multiparty computation could achieve <1. Despite limitations in severe fault detection and dataset diversity, AF2D demonstrates robust performance. Future work includes integrating multi-modal data (SCADA, vibration, environmental), testing real-time deployment, and expanding federated datasets to enhance generalizability and scalability.OPEN ACCESS Received: 11/09/2025 Accepted: 16/10/2025</p>

== Document ==
<pdf>Media:Draft_Sanchez Pinedo_521166662-8340-document.pdf</pdf>

← Older revision	Revision as of 09:18, 19 December 2025
(No difference)

← Older revision	Revision as of 08:50, 19 December 2025
(No difference)

Irfan et al 2025a - Revision history

Scipediacontent: Scipediacontent moved page Review 538456560717 to Irfan et al 2025a

JSanchez: JSanchez moved page Draft Sanchez Pinedo 521166662 to Review 538456560717

JSanchez: Created page with " == Abstract == Wind turbine reliability is critical for sustainable energy production, yet fault diagnosis faces challenges due to data privacy concerns, heterogeneous op..."

JSanchez: Created page with " == Abstract ==
Wind turbine reliability is critical for sustainable energy production, yet fault diagnosis faces challenges due to data privacy concerns, heterogeneous op..."