Abstract

Sometimes, smart heating control applications are partially equipped with missing values and outliers in the sensor data due to software/hardware failures/human errors. To provide an effective analysis and decision-making, erroneous sensor data should be cleaned by imputation of missing values and smoothing outliers. In this paper, we present a case of the Smart Heating Control System (SHCS) installed in the South Ural State University, and describe the structure and development principles of Data Cleaning Module (DCM) of the system. We implement DCM through data mining and neural network technologies as a set of the following subsystems. The preprocessor extracts raw data from the system's data warehouse and prepares a training data for further processing. Predictor provides Recurrent Neural Network (RNN) to forecast the next value of a sensor based on its historical data. Reconstructor determines if the current value of a sensor is an outlier, and if so, imputes it by the synthetic value from Predictor. Finally, Anomaly Detector subsystem discovers anomalous sequences in the sensor data. In the experiments on the real sensor data, DCM showed relatively high and stable accuracy as well as adequate detection of anomalies.