Diffusion-based digital twin-driven adversarial domain adaptation for fault diagnosis in high-energy beam choppers

As critical rotating machinery for regulating high-energy particle beams, beam choppers play a crucial role in ensuring the stable operation of high-energy scientific facilities through fault diagnosis. Traditional fault diagnostics often assume a consistent distribution between training and testing data and rely on sufficient samples to train reliable diagnostic models. However, these assumptions are often impractical because high-energy beam choppers operate under different conditions. This leads to distribution shifts that degrade the accuracy and reliability of diagnostics. For this reason, a novel digital twin-driven adversarial domain adaptation (DTADA) based on a diffusion model is proposed. Specifically, a convolutional autoencoder is first trained solely using normal data to extract low-dimensional features from vibration signals. The extracted features are then used in a digital twin-driven diffusion model, which first gradually changes the data to pure noise and then learns the denoising process to generate a synthetic twin similar to the real data. By assigning real-world data to the target domain and generated twin data to the source domain, an improved adversarial domain adaptation is developed using Wasserstein distance and gradient penalty to enhance feature differentiation and distribution alignment. The proposed DTADA was evaluated through fault diagnosis experiments of the beam chopper. Results demonstrate that the proposal achieves high diagnostic performance with misaligned data distribution as well as insufficient measured data. It offers significant advantages for the fault diagnosis of high-energy beam choppers.