This paper presents a methodology for data cleaning of sensor data using the Kalman filter. The Kalman filter is an on-line algorithm and as such is ideal for usage on the sensor data streams. The Kalman filter learns parameters of a user-specified underlying model which models the phenomena the sensor is measuring. Usage of the Kalman filter is proposed to predict the expected values of the measuring process in the near future and to detect the anomalies in the data stream. Furthermore the Kalman filter prediction can be used to replace missing or invalid values in the data stream. Algorithm only requires sensor measurements as an input, which makes it ideal to be placed as near to the resource tier in the N-tier architecture as possible.