Kalman Filter and Artificial Neural Network for Real time Sensor Denoising

Obtaining more accurate and precise information from real-time data requires highly precise sensor readings. The priori and the fundamental step before proceeding with further data analysis and taking conclusions is the denoising of sensor data. Considering this requirement, this article assesses the effectiveness of Kalman filters and feed forward artificial neural networks for the purpose of sensor denoising. To accomplish this objective, we designed an Internet of Things (IoT) device employing a Raspberry Pi 4 and a digital humidity and temperature sensor (DHT 11). After the prototype development, we deployed a discrete time Kalman filter, integrated with ThingSpeak, and then we collected data. The data was filtered using two standard deviations. After that, we performed the outlier's removal and Normality test. Finally, we implemented the artificial neural network to train and test the algorithm, using feed-forward artificial neural network (FF - ANN) with a sequential architecture. For inference, we evaluated the Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Coefficient of Determination (R2 score), Central Processing Unit Time (CPU.T), Central Processing Unit memory (CPU) and - Standard Deviation (StDev). The results indicated that the Kalman filter is the optimal solution for sensor denoising, reducing data fluctuations, and removing temperature and humidity outliers in real-time for online scenarios. However, it should be noted that FF-ANN exhibits superior MAE, MSE, RMSE, and R2 Score because it was implemented on preprocessed data. The findings showed that the Kalman filter has 3.5 speeder than FF-ANN on convergence time.