Recurrent Neural Network Based Modelling of Industrial Grinding Time Series Data

Modelling the time series data generated from a complex and nonlinear industrial grinding unit mandates the use of sophisticated algorithms capable of efficiently approximating the system under consideration. Recurrent Neural Networks, which are proven to be competent enough to approximate many time series systems, can be utilized for identification of industrial grinding circuits. However, the usage of RNNs for system identification tool is limited due to the heuristic estimation of network hyper parameters viz., number of hidden layers to be explored, number of nodes in each hidden layer, activation function and number of previous time instances to be considered for capturing the dynamics of the process. In this study, we address this heuristic approach by proposing an algorithm which can determine the optimal values of these hyper parameters for RNNs. This optimal determination of hyper parameters is done by adopting a multi-objective optimization problem with maximization of the accuracy of the developed model and minimization of the number of nodes in the network as the two conflicting objectives. The performance of the proposed algorithm on a real life industrial grinding circuit data shows its success and competitiveness.