Deep auto-encoded Conditional Gaussian Mixture Model for warranty claims forecasting

Forecasting warranty claims enables (i) identifying and rectifying quality and reliability problems at the early stages of production, (ii) improving future product designs, and (iii) efficiently allocating financial resources for warranty and maintenance. Nevertheless, precise forecasting of warranty claims is challenging due to the data maturation phenomenon, which is characterized by challenges that include market demand variability, reporting delays, heterogeneous production quality, and customer late claim rush. This paper proposes the Deep auto-encoded Conditional Gaussian Mixture Model (DCGMM) for warranty claims forecasting. DCGMM is a novel hybrid Bayesian and deep learning framework. The method learns a prior joint distribution between auto-encoded temporal latent embeddings extracted from early observed cumulative warranty claims and long-term mature warranty claims based on historical products. DCGMM then uses Bayesian inference to forecast the mature warranty claims of an in-service product of interest conditioned on its auto-encoded embeddings extracted from its limited early claims observations. DCGMM is robust to product variability and complex temporal trends because of its ability to identify different cluster or sub-population behaviors. Furthermore, learning a lower-dimensional space is essential to achieving tractable Bayesian inferences. We validate the performance of the DCGMM for warranty claims forecasting on a large-scale automotive case study and compare it to other state-of-the-art time-series forecasting algorithms.