Attention-based fuzzy neural networks for self-supervised data annotation

Annotating vibration data from heavy-duty pumps in the mining industry is highly challenging because it demands domain knowledge, a complex inspection setup, and, in many cases, remains infeasible. A self-supervised data annotation (SSDA) framework is therefore proposed and evaluated on historical data of slurry-pump vibration signals. The framework began with the collection of heterogeneous information, followed by information fusion using an autoencoder. This was then followed by a datafication step for preprocessing and achieving a better representation of features through a feature embedding technique. As a result, redundant information was pushed into an eight-dimensional latent space, achieving a reconstruction loss of 0.0023. Furthermore, Initial data annotation was obtained by combining the Isolation Forest and Kneedle algorithms to locate a data-driven knee or threshold, and it was found to be 0.58 for predicting labels. Partial samples were labeled and considered accurate. Lastly, an attention-based fuzzy neural network (AFNN) is trained on those labels where membership functions convert each latent feature into graded truth values. At the same time, an attention layer highlights the most relevant rules. An iterative self-training loop was implemented to refine the training set and obtain labeled data with higher model confidence. Here, we also tested six baseline models and found AFNN quite impressive. After seven iterations 2780 of 2872 samples were labeled and the remaining 92 are considered uncertain, still need some review from an expert, and the AFNN model confidence was (96.8%). Statistical analysis confirmed that the model predictions were significantly associated with true labels ( p < 0 . 05 ) and not driven by chance. • A self-supervised data annotation framework is proposed for industrial slurry-pump vibration signals. • Sensor features are fused and compressed into an eight-dimensional latent space with a reconstruction loss of 0.0023. • Initial labels are generated using Isolation Forest and a data-driven Kneedle threshold (0.58). • An attention-based fuzzy neural network refines annotations through iterative self-training. • The framework annotates 2780 of 2872 samples (96.8% coverage) with high confidence and outperforms six baseline models.