SVD-based Particulate Matter Estimation Using LSTM-Based Post-Processing for Collaborative Virtual Sensor Systems

Research on particulate matter digital twinning spans product manufacturing processes to individual health. To obtain particulate matter, we acquire particle count from raw data and then apply corrections using transfer function. Research have been conducted to replicate the transfer function of a high-performance device using singular value decomposition with a low-cost, low-power device. However, this replicated transfer function retains noise components. This paper proposes using LSTM for post-processing, achieving smoother signals and noise reduction. The experimental results show that post-processing with LSTM yields significantly lower root-mean-square error (2.1692) when compared to other filters: mean filter (3.4681), low-pass filter (3.5828), and Kalman filter (3.3866).