The proliferation of IoT sensors and advanced monitoring systems creates unprecedented opportunities to develop digital twins across various industrial sectors and diverse engineering applications. This paper extends previous work [1] on building scalable, data-driven digital twins that can be efficiently trained and deployed across individual assets or fleets for multi-channel heterogeneous sensor inputs. Unlike traditional approaches focusing on dimensional reduction for computational efficiency, our methodology emphasises learning interpretable latent variable models that disentangle different sources of variance in the sensor data. Multi-modal sensor data are transformed into latent spaces, prioritising untangled interpretable variance sources over mere efficient representation [2]. This approach prioritises the digital twin to capture and model distinct physical phenomena, operational modes, and environmental factors influencing asset behaviour. The resulting latent representations facilitate a better understanding of the underlying system dynamics and enable more targeted interventions for real-world applications. The proposed architecture supports hybrid digital twin models that combine data-driven learning with physics-based knowledge, though this flexibility comes at the cost of reduced scalability. We demonstrate the framework's effectiveness on simulated datasets for different industrial settings and sensor configurations, demonstrating that our approach can adapt to varying sensor configurations, operational conditions, and application requirements while prioritising the interpretability of the learned representations. Key contributions include i) a scalable architecture for training and deploying digital twins across heterogeneous assets and applications, ii) novel approaches to latent space transformation that prioritise interpretability over compressibility, and iii) a preliminary crafted study demonstrating that our framework provides a promising foundation for developing and deploying more interpretable and adaptable digital twins at scale.