A federated deep domain adaptation‐based framework for nuclear power steam turbines considering privacy‐preserving

Abstract With the increasing awareness of environmental protection, sustainable manufacturing has become an important component in various industries. As an essential foundation for sustainable strategy, safe and reliable operation and maintenance of nuclear power resources is crucial, requesting agile and precise response and diagnosis of equipment failure signals. Due to security requirements, nuclear power plants strictly isolate operating data and form an actual data island. Simultaneously, the insufficient amount of fault sample data makes it difficult to establish an accurate fault diagnosis model. How to establish a stable and reliable nuclear power steam turbine vibration fault diagnosis model across different nuclear power plants and nuclear equipment has become a big problem. To achieve secure model aggregation without violating client privacy, federated learning (FL) has become a research hot spot for model aggregation, but it ignores the differences between source clients and fails to capture domain‐invariant features during local training, which hinders its further development. To address this challenge, a federated deep domain adaptation‐based framework considering privacy‐preserving (FL‐DDA) is proposed for operations and maintenance in nuclear power plants. The framework performs feature extraction locally in source nuclear power plants and targets nuclear power plants, such that the features are shared securely without revealing data privacy. At the same time, domain adversarial training is integrated into the local model training to realise the transfer of vibration fault diagnosis knowledge. Furthermore, an adaptive weight mechanism is devised to facilitate the adaptive adjustment of model weights in federated aggregation. Finally, a desensitised vibration dataset in nuclear power steam turbines is applied for validation, and FL‐DDA is compared with other existing methods. Under the premise of data privacy security, the proposed FL‐DDA framework proves to outperform its peers in vibration fault diagnosis and domain adaptation.