Integrating physical knowledge for generative-based zero-shot learning models in process fault diagnosis

Despite longstanding operational processes, persistent undiagnosed issues and inefficiencies continue to exist. Specifically, the collected data with specified faults is often either non-existent or sparse. The challenge lies in the absence of specified faults for reliable training of fault diagnosis models in such processes. This study proposes an innovative physical knowledge-guided zero-sample fault diagnosis method, which decomposes process variables into states and defines attributes based on domain knowledge. This transformation of variables into the attribute space replaces the traditional label space. The method involves three key steps: (1) Constructing the Seen Fault Latent Space: Utilizing seen fault data through the Conditional Variational Auto-Encoder model and Linear Discriminant Analysis in the latent space to classify seen faults. (2) Extending the Model Space with Unseen Fault Attributes: Using attributes to extend the unknown fault space and introducing a discriminator to ensure accurate separation of seen and unseen faults. (3) Retraining the Model: Using the data generated in the second step to retrain the model, enabling the diagnosis of both seen and unseen faults by the encoder. Experiments on numerical, continuous stirred tank reactor, and three-level tank examples demonstrate a significant 11 % improvement in classification accuracy for unseen fault samples compared to traditional methods.