Conditional variational learning for collaborative wind turbine diagnostics and out-of-distribution detection

The normal operation of wind turbine systems is critical for the stability and safety of energy production. Timely fault detection is essential to prevent system failures, especially given the challenges posed by diverse operational conditions and unseen domains. In real-world scenarios, the presence of unseen wind turbines, or domains without historical data, significantly impacts the performance of traditional models. This paper proposes a collaborative domain-invariant approach for generalized fault detection, combining domain generalization (DG) and domain-level out-of-distribution (OOD) detection. This unified approach addresses the critical gap where existing methods treat cross-domain generalization and safety-critical anomaly identification as separate problems. The model categorizes input data into environmental and operational attributes, leveraging a disentangled latent space. Regularization techniques are applied to facilitate subspace disentanglement and ensure invariant learning, which enhances the model ability to generalize across unseen turbines. Additionally, normalizing flows are integrated for effective OOD detection. The effectiveness of the proposed method is validated through two case studies from real wind farms to provide comprehensive evaluation. The leave-one-domain-out evaluation is conducted iteratively, demonstrating the superior performance of the method in cross-turbine fault diagnosis and OOD detection. An extensive ablation analysis further highlights the advantages of the proposed approach, including reduced inter-turbine accuracy variance and enhanced OOD detection. The method outperforms state-of-the-art models in both fault diagnosis and OOD detection. Experimental results show that the model is robust across multiple turbine domains, offering a promising solution for wind turbine fault detection in dynamic environments. © 2026 Elsevier Ltd.