A key challenge in real-world structural health monitoring (SHM) is diversity of damage phenomena and variability in environmental and operational conditions. Conventional learning techniques, while adequate for moderately complex inference tasks, can be limiting in highly complex and rapidly changing environments, especially when insufficient data is available. We present an adaptive learning methodology where stochastic models continuously evolve with the time-varying environment and Dirichlet process mixture models are utilized to self-adapt to structure within the data. Coupled with appropriate physics-based phenomenology, the approach provides an adaptive and effective framework for online SHM. The proposed technique is demonstrated for the detection of progressive fatigue damage in a metallic structure under variable-amplitude loading.