Big data stream learning based on hybridized Kalman filter and backpropagation through time method

Most real-time control systems are often accompanied with various changes such as variations of working load and changes of the environment. Hence it is necessary to perform real-time process modeling so that the model can adjust itself in runtime to maintain high accuracy of states under control. This paper considers process model represented as a deep recurrent neural network. We propose a new hybridized learning method for online updating the weights of such recurrent neural networks by exploiting both fast convergence of Kalman filter and stable search of the Backpropagation through time algorithm. Several experiments were made to show that the proposed learning method has fast convergence, high accuracy and good adaptivity. It can not only achieve high modeling accuracy for a static process but also quickly adapt to changes of characteristics in a time-varying process.