Neural networks for equipment failure prediction based on unstructured production data

Industrial enterprises accumulate large volumes of unstructured data from sensors, event logs, and technical reports; however, their potential for failure prediction often remains underutilised. The aim of this study was to demonstrate how modern neural networks transform such heterogeneous data streams into early failure warnings and how the integration of these models into maintenance systems reduces downtime and increases equipment reliability. The methodology combined a unified data preparation pipeline – including cleaning, timestamp matching, and event-based and semantic feature extraction from report texts – with a comparative evaluation of three classes of architectures: convolutional networks for vibration signals and images, recurrent networks capable of modelling long-term dependencies in time series, and attention-based models for multimodal learning. The main results focused on forecast quality and operational impact. Models based on attention mechanisms consistently outperformed other approaches in terms of weighted average accuracy by 5-9 percentage points, providing a warning horizon of 24-72 hours. Convolutional networks demonstrated the highest sensitivity to high-frequency vibration patterns, while recurrent models were more effective in tracking slow degradation trends. Integration of the best-performing model into a maintenance information system – with automated request generation and work prioritisation – reduced downtime by 14-21%, increased mean time to failure by 10-15%, and decreased the proportion of false alarms to a level acceptable to dispatchers. Available production cases were summarised, and theoretical application scenarios were outlined for industries in which empirical data remain limited. The practical significance of the study lies in a reproducible approach to processing and interpreting unstructured data, as well as in the demonstrated contribution to reducing downtime and increasing the reliability of the equipment fleet