APPLICATION OF ARTIFICIAL NEURAL NETWORK IN PAPER INDUSTRY

Abstract

For consistency control one case is a SISO system for which both Lambda and Ziegler Nichols( Z-N )tuning methodologies have been used to find out the PID and PI control parameters for both analog(S-domain) and digital( Z-domain) control system. As usual for Z-N tuning, however Bode's stability criterion has been used. After developing the characteristic models, a PI/PID based control loop and the corresponding SIMULINK model have been developed. For the same case of SISO consistency control an ANN control system (BPNN) along with the necessary algorithm has also been designed. The network is trained with PI/PID simulated data. Similarly the second case of consistency control with two inputs and single output (TISO) has been dealt with assuming negligible interactions between parameters. Accordingly simulations have been carried out for both PI and ANN controllers. For stock flow control model equations are already available. In this case mill data for training for ANN controller are employed. Data predicted from theoretical models are also used for training purposes. For the case of total head control an example of control loop with hydraulic headbox available in literature has been used for analysis. The dynamic models of all the elements of closed loop have been found out and closed loop transfer function are developed. This has been used for further analysis through MATLAB simulation and neural computation. Similar analyses have been made for stock level, pH, temperature and basis weight, first by developing appropriate dynamic models in laplace domain, converting to Z-domain, then designing classical control loops and analyzing them. These are then followed by transforming in to a neural network based control system. It is well known that if the process interactions are significant, even the best multi-loop control system may not provide satisfactory control. In these situations there are iv incentives for considering multivariable control strategies such as decoupling control and model predictive control. Hence, development of control system for MIMO systems with some examples from paper machine wet end including headbox has been attempted. The examples are: interactions of total head and stock level, air pressure and stock level (for air cushion head box), retention on forming wire and consistency control with two inputs and single output (TISO)system. For multi input multi output (MIMO) system considered in this present investigation both relative gain array (RGA) and decoupling techniques are used. For the case of MIMO system, however, the same procedures as in the case of SISO system have been followed. The only additional parameters of control included in the analysis of MIMO system for estimating the degrees of interaction and pairing of controlled and manipulated variables between different sets of control loops have been the relative gain array method (RGA) and decoupling technique for adjusting the interaction. The relative gain between 0 to 1 are only considered for analysis. In chaOter'-6, an attempt has been made to compute data from various models for both SISO and MIMO system using the classical controller and neural network based control with the help of MATLAB SIMULINK toolbox. The procedures laid down in Chapter-4, the various equations presented therein, the algorithm developed for the ANN and for PID, and finally the models developed for the various wet end parameters given in Chapter-5 are used. From the plethora of data from MATLAB Simulation of the process parameters, some dynamic characteristics have been drawn in various graphs with response as a function of time for all the above mentioned parameters when unit step input signal is applied as a forcing function and then performances are evaluated. While for consistency control, the dynamic responses using both PI & PID are studied and compared with performances of ANN based controller, the cases of total head, stock level and pH only PI and ANN, for temperature and basis weight only PID and ANN are employed, analyzed and compared. Conclusions based on the present study are finally drawn in the concluding Chapter-7, conclusions and recommendations. Recommendations based on the present work, limitations and scope of further study are also briefly discussed in this concluding chapter