Modern industrial assets operate under diverse and changing working conditions, such as varying loads, temperatures, production rates and environmental stresses. Meanwhile, advances in sensing, monitoring and industrial data collection provide increasingly rich information on failures, degradation and operating environments. The central challenge is how to convert these heterogeneous data into effective and economically justified maintenance actions. This talk presents two complementary approaches to data-driven maintenance management under varying working conditions. The first considers preventive maintenance when historical failure-time data and working-condition features are available, but the underlying lifetime distribution is unknown. The second addresses condition-based maintenance under dynamic working conditions and component heterogeneity. Together, the two approaches illustrate a broader transition from model-centred maintenance analysis towards decision-centred learning.