Abstract
Machinery failure prognostic, aiming to predict remaining useful life (RUL) of equipment by a series of measures, is rarely applied in the mineral processing industry. The commonly used algorithms ignore the spatial heterogeneity and temporal characteristics of equipment degradation, affecting the prediction accuracy and uncertainty quantification. This paper proposes an advanced uncertainty-aware machinery failure prognostic method for mineral manufacturing equipment. The algorithm adopts Monotonic Recurrent Autoencoder to extract the valuable degradation information from complex real-world data. Building on this, a Weighted Sequence Neural Processes (WSNP) framework is developed, consisting of encoding and decoding modules. The encoder simultaneously processes the spatial heterogeneity and temporal dynamics of degradation data, enabling more accurate inference of uncertainty associated with target degradation. The decoder enables precise inference of the RUL distribution by leveraging the extracted degradation information and uncertainty encoding. Bayesian variational is applied to WSNP inference process, circumventing the complex analytical process of the marginal likelihood integral. The proposed algorithm is applied to a real-world mineral manufacturing equipment, and the experimental results confirmed its effectiveness and feasibility.