Abstract
Traditional deep-learning-based approaches often struggle with data imbalance and variability across fault conditions and normal scenarios, especially in industrial processes. Besides, inconsistent feature distributions from combining different fault conditions into the same category are a limitation for many data-driven algorithms. This study proposes a fault detection framework that combines Siamese neural networks with transfer learning, using a pretrained fault diagnosis model as its backbone, taking advantage of knowledge related to the attribute space that characterizes individual fault patterns. Our method transforms the detection classification problem into an embedding similarity task, allowing for improved differentiation between normal and faulty operations. This approach poses an alternative for data imbalance and a lack of labeled anomaly data, as it is based on the combination of normal and faulty time series. Our best model achieved an F1-score of 91.41% on the test set, and the t-distributed stochastic neighbor embedding indicates that the knowledge transferred from diagnosis allowed the detection model to generate embeddings that discriminate between most faulty conditions. When analyzing individual fault detection rates, we observed that our model demonstrated superior performance compared with recent literature for most fault cases.