Font Size: 

We consider time series of field failure data (warranty claims) of domestic appliances, manufactured by different plants, with the aim of forecasting failures within the warranty period. The failure profiles over two-year periods display variation across monitoring epochs and also batch-to-batch variation. A non-linear state space model is developed to jointly represent the variation of the underlying failure rate parameters and the observed occurrence of failures, obtaining a dynamic Poisson-Lognormal model with a meaningful covariance structure of failure rates between monitoring epochs. An adaptation of the auxiliary particle filter similar to the particle learning algorithm is used for parameter learning and forecasting. A series of examples with data from different production plants show that it is possible to obtain a small forecasting error for claims having very different patterns.