Warranty services are essential for managing product reliability but impose substantial financial burdens due to failures during the warranty period. Traditional lifetime distribution-based methods, such as Weibull and log-normal models, have been wid...
Warranty services are essential for managing product reliability but impose substantial financial burdens due to failures during the warranty period. Traditional lifetime distribution-based methods, such as Weibull and log-normal models, have been widely used for field reliability prediction, however, their accuracy has been reported to deteriorate when only limited early warranty data are available.
This study proposes a Weighted Long Short-Term Memory (WLSTM) model for long-term failure prediction. The model integrates sequential inputs with periodic auxiliary variables, and employs a weighted loss function with failure-occurrence and time-dependent weights to mitigate the effects of data sparsity and imbalance. Furthermore, to capture heterogeneous failure-rate patterns, clustering is applied, enabling group-specific learning that enhances prediction performance.
Through this methodology, monthly failure counts are forecasted and subsequently converted into the cumulative failure probability , which serves as the primary reliability measure. Empirical validation on large-scale warranty data from twelve categories of home appliances demonstrates that the proposed model achieves superior performance compared with conventional statistical distributions and advanced deep learning methods across multiple clusters. These findings confirm that the proposed approach enables accurate and stable long-term failure prediction, even when only limited early data are available.