The optimization of asset management policies for engineering systems can be characterized as a stochastic, sequential decision problem. Crucially, it relies on an accurate representation of the underlying real-world processes. A central bottleneck in maintenance modeling is the scarcity of informative asset-level failure and repair data. Typically, only a handful of failure and repair events are observed per component over several years of operation, and these observations are insufficient to reliably infer degradation behavior. However, for many assets, a variety of large-population-based statistics are available, for example, average failure rates or failure-time distributions. In this contribution, we propose a method to merge asset-specific data with population-wide statistical evidence to allow efficient learning of a generative model of latent degradation states and observable system responses, even in data-poor regimes. We jointly model asset degradation and measured signals through a partially observed Markov process. Population-level failure statistics are incorporated as additional data constraining the latent parameters, and combined with observations into a coupled Bayesian inference framework. While the statistics guide and regularize the latent dynamics, observation sequences provide information about the system response at different degradation stages. After training, we embed the model in a decision framework as a partially observable Markov decision process (POMDP) and compute maintenance policies by solving the resulting POMDP using reinforcement learning. The proposed approach enables data-driven degradation modeling and decision optimization, even when data is scarce and only a few observations of interest are available. Learning is aided by including population-wide statistics into the inference pipeline, and the resulting model is both consistent with asset-specific responses and with population-wide statistical properties.