Applicability of Neural Networks for Processing Low-Dimensional Sensor Data on Embedded Systems: Case Study

This research proposes a new noise-resistant approach for predicting water levels using the capacitive sensor output. The method was tested on data from different sensor configurations. The method performance was compared to a neural network-based (NN) model. General NNs are often used for data linearization and noise compensation in smart sensor systems. However, the domain-aware approach is shown to be better suited for usage on embedded systems in terms of accuracy and memory consumption. The testing on datasets from three different sensor configurations shows 1.3x to 7.7x lower RMSE than the NN models. The proposed model also has fewer parameters, requiring 6x to 9x less memory. This result was validated on a real-time Arm Cortex-M0 embedded device.