A data-driven approach to estimate flow fields from sparse distributed sensors in negative pressure wards

Growing concerns about air quality in medical environments and its health implications have highlighted the need for fast and accurate reconstruction of airflow patterns in complex indoor settings to support monitoring and pollutant assessment. To address this challenge, a data-driven approach integrating offline numerical simulations with online sparse measurements is employed to reconstruct airflow patterns efficiently and accurately in a negative pressure isolation ward. Firstly, dimensionality reduction strategies are evaluated by comparing both linear and nonlinear methods. Numerical simulations under 45 typical conditions are processed using proper orthogonal decomposition (POD) for dimensionality reduction, reducing computational effort by over 80 %. Although convolutional autoencoders (CAE) better capture nonlinear features, POD is adopted here for its high efficiency and physically interpretable modes, offering clear advantages for fast reconstruction tasks in engineering scenarios. To reconstruct missing flow fields, two sensor placement strategies—random iterative and spectral clustering—are compared, achieving reconstruction errors below 5 % on typical cross-sections. Fast prediction of flow fields is further enabled by sparse online data, with optimal sensor placement analysis determining 15 as a stable number. Additionally, Three-dimensional reconstruction captures both global patterns and local details, with overall root mean square errors within 5 %, validating the method's accuracy. This approach reduces sensor deployment demands while maintaining high reconstruction quality, offering practical support for airflow optimization and infection risk assessment in medical environments. The comparative analysis with CAE also provides a foundation for incorporating nonlinear methods in future research.