Performance analysis of multi-layer perceptron regression model with mixed-rate sensor data inputs

Abstract

Multi-dimensional sensory data would likely have uneven resolution, hence, necessitating pre-processing steps before these could be useful for data analytics. In this study, we analyzed the performance of a multivariate regression model using multi-layer perceptron subjected to inputs that have asynchronous sampling rates. We adopted a publicly available sensor-based dataset to explore the feasibility of the MLP for the predictive analytics of data gathered at mixed-sampling rates. The sensor data consisted of temperature and humidity measurements at 1.67, 1.11 and 0.55-mHz sampling frequencies. Applying MinMax normalization to the input layer, the RMSE of the regression model with Gradient Descent optimizer was reduced by around 8 per cent. Downsampling the input data using spatial filters improved the performance of the MLP regression model based on its generally lower RMSE and higher R² values compared to the baseline. Our findings suggest that MLP networks provide a suitable baseline analytics on mixed-rate input data. Intricate methods could later build on the architecture presented in our current study.